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): single source of truth for a run's primary engine (fixes mid-run scope_mismatch crash) (#4745)

LearningCircuit8f5d57aa47bd
diff --git a/src/local_deep_research/advanced_search_system/strategies/langgraph_agent_strategy.py b/src/local_deep_research/advanced_search_system/strategies/langgraph_agent_strategy.py
index ebebd3b4f2..5692c0addd 100644
--- a/src/local_deep_research/advanced_search_system/strategies/langgraph_agent_strategy.py
+++ b/src/local_deep_research/advanced_search_system/strategies/langgraph_agent_strategy.py
@@ -19,11 +19,9 @@ from langchain_core.tools import tool
from langgraph.errors import GraphRecursionError
from loguru import logger
-from ...utilities.type_utils import unwrap_setting
from ...citation_handler import CitationHandler
from ..tools.fetch import FETCH_MODES, build_fetch_tool
from .base_strategy import BaseSearchStrategy
-from ...constants import DEFAULT_SEARCH_TOOL
from ...security import sanitize_error_for_client
# ---------------------------------------------------------------------------
@@ -574,15 +572,23 @@ class LangGraphAgentStrategy(BaseSearchStrategy):
from local_deep_research.security.egress.policy import (
PolicyDeniedError,
context_from_snapshot,
+ resolve_run_primary_engine,
)
- primary = self._get_current_engine_name() or unwrap_setting(
- self.settings_snapshot.get("search.tool")
- )
try:
- return context_from_snapshot(
- self.settings_snapshot, primary or DEFAULT_SEARCH_TOOL
- )
+ # Derive the primary engine the SAME way the factory PEP does —
+ # from ``search.tool`` — NOT from the engine class name. Under the
+ # default ADAPTIVE scope the primary IS what resolves the concrete
+ # scope, so a divergent primary here silently under-filters the
+ # agent's tool list: a private collection primary classified via the
+ # class heuristic ("libraryrag" -> unknown -> BOTH) left public
+ # engines visible, which the factory then hard-denied mid-run
+ # (scope_mismatch_private_only). resolve_run_primary_engine raises
+ # ValueError when no primary is configured; this advisory filter
+ # then degrades to unfiltered (the factory PEP still enforces) —
+ # research_service has already failed the run closed by that point.
+ primary = resolve_run_primary_engine(self.settings_snapshot)
+ return context_from_snapshot(self.settings_snapshot, primary)
except PolicyDeniedError:
# Corrupted/invalid policy.egress_scope — re-raise so the
# caller fails closed instead of silently running unfiltered.
diff --git a/src/local_deep_research/security/egress/policy.py b/src/local_deep_research/security/egress/policy.py
index 9ca95c55b0..4d24c6091a 100644
--- a/src/local_deep_research/security/egress/policy.py
+++ b/src/local_deep_research/security/egress/policy.py
@@ -28,7 +28,6 @@ from loguru import logger
from ..network_utils import is_private_ip
from ...utilities.type_utils import unwrap_setting
-from ...constants import DEFAULT_SEARCH_TOOL
# Bounded DNS lookup timeout; getaddrinfo has no native timeout kwarg, so
@@ -541,12 +540,7 @@ def filter_candidates_by_egress(
"adaptive",
):
return engine_names
- primary = (
- _get_setting_value(
- settings_snapshot, "search.tool", DEFAULT_SEARCH_TOOL
- )
- or DEFAULT_SEARCH_TOOL
- )
+ primary = resolve_run_primary_engine(settings_snapshot)
ctx = context_from_snapshot(
settings_snapshot,
primary,
@@ -1330,6 +1324,63 @@ def _get_setting_value(snapshot: dict, key: str, default):
return unwrap_setting(raw)
+def resolve_run_primary_engine(
+ settings_snapshot: Optional[dict], default: Optional[str] = None
+) -> str:
+ """Resolve a research run's primary search engine id from a snapshot.
+
+ Single source of truth for "which engine drives this run's egress
+ scope". Under the default ADAPTIVE scope the concrete scope is derived by
+ classifying THIS primary, so run-scoped callers that build an
+ ``EgressContext`` should derive the primary here. Otherwise two layers can
+ resolve ADAPTIVE to different scopes and one silently under-enforces — the
+ LangGraph tool-list filter once derived the primary from the engine CLASS
+ name (a collection -> ``"libraryrag"`` -> unclassified -> BOTH) while the
+ factory PEP derived it from ``search.tool`` (the real collection key ->
+ PRIVATE_ONLY), so a public engine reached the agent and was then
+ hard-denied mid-run (``scope_mismatch_private_only``).
+
+ Reads ``search.tool`` (flat or ``{"value": ...}`` shaped). A blank-after-
+ strip (``" "``) or non-string (``5``, list/dict) value is treated as
+ MISSING — it must not slip through the truthiness check and classify to the
+ permissive BOTH scope.
+
+ A run with NO configured primary is a wiring error, not a normal state.
+ Silently substituting a default would set the egress scope from an engine
+ the user never chose — and since the system default is public (``searxng``),
+ that fails OPEN. So a missing/empty ``search.tool`` raises ``ValueError``
+ unless the caller passes an explicit ``default``:
+
+ * run-level callers (research_service, the strategy tool-list filter,
+ ``filter_candidates_by_egress``) pass none and fail closed — the worker
+ refuses the run; advisory filters degrade to unfiltered with the factory
+ PEP still enforcing.
+ * the search-engine factory passes ``default=engine_name`` because it is
+ evaluating that one specific engine, so deriving the scope from it is
+ meaningful rather than arbitrary.
+
+ Raises:
+ ValueError: ``search.tool`` is missing/blank/non-string and no truthy
+ ``default`` is given.
+ """
+ primary = (
+ _get_setting_value(settings_snapshot, "search.tool", None)
+ if settings_snapshot
+ else None
+ )
+ # Blank-after-strip or non-string is not a usable engine id → missing.
+ primary = primary.strip() if isinstance(primary, str) else None
+ if primary:
+ return primary
+ if default:
+ return default
+ raise ValueError(
+ "no primary search engine configured: settings 'search.tool' is "
+ "missing, blank, or not a string, so the run's egress scope cannot "
+ "be resolved"
+ )
+
+
def _coerce_bool(value) -> bool:
"""Coerce a setting value to bool, defensively.
diff --git a/src/local_deep_research/web/queue/processor_v2.py b/src/local_deep_research/web/queue/processor_v2.py
index 5754984f87..af1767bae2 100644
--- a/src/local_deep_research/web/queue/processor_v2.py
+++ b/src/local_deep_research/web/queue/processor_v2.py
@@ -1171,7 +1171,15 @@ class QueueProcessorV2:
complete_settings = settings_snapshot.get("settings_snapshot", {})
else:
submission_params = settings_snapshot
- complete_settings = {}
+ # A legacy-flat queued row (enqueued by an older version, pre-
+ # "submission" wrapper) carries no settings_snapshot. Seed the run's
+ # primary engine from the submitted search_engine so the worker's
+ # egress build (resolve_run_primary_engine) doesn't fail closed on
+ # an empty snapshot and refuse the run.
+ _legacy_engine = submission_params.get("search_engine")
+ complete_settings = (
+ {"search.tool": _legacy_engine} if _legacy_engine else {}
+ )
try:
research_thread = start_research_process(
diff --git a/src/local_deep_research/web/routes/research_routes.py b/src/local_deep_research/web/routes/research_routes.py
index 61a287d330..69b7056179 100644
--- a/src/local_deep_research/web/routes/research_routes.py
+++ b/src/local_deep_research/web/routes/research_routes.py
@@ -231,6 +231,7 @@ def _precheck_engine_policy(settings_manager, params, search_engine, username):
PolicyDeniedError,
context_from_snapshot,
evaluate_engine,
+ resolve_run_primary_engine,
)
policy_snapshot = settings_manager.get_settings_snapshot()
@@ -245,13 +246,16 @@ def _precheck_engine_policy(settings_manager, params, search_engine, username):
# a global settings save).
_apply_policy_overrides(policy_snapshot, params)
- primary = settings_manager.get_setting(
- "search.tool", DEFAULT_SEARCH_TOOL
- )
+ # Resolve the primary the SAME way the worker does (single source of
+ # truth) so the precheck and the background worker agree on accept vs.
+ # refuse — including the fail-closed missing-primary case, which the
+ # ValueError handler below maps to a 400. (Previously this substituted
+ # searxng and accepted runs the worker then refused.)
try:
+ primary = resolve_run_primary_engine(policy_snapshot)
policy_ctx = context_from_snapshot(
policy_snapshot,
- primary or DEFAULT_SEARCH_TOOL,
+ primary,
username=username,
)
except PolicyDeniedError as exc:
diff --git a/src/local_deep_research/web/services/research_service.py b/src/local_deep_research/web/services/research_service.py
index bc4df9b0bc..032c877841 100644
--- a/src/local_deep_research/web/services/research_service.py
+++ b/src/local_deep_research/web/services/research_service.py
@@ -34,7 +34,6 @@ from ...utilities.threading_utils import thread_context, thread_with_app_context
from ..models.database import calculate_duration
from ...settings.env_registry import get_env_setting
from .socket_service import SocketIOService
-from ...constants import DEFAULT_SEARCH_TOOL
OUTPUT_DIR = get_research_outputs_directory()
@@ -941,20 +940,14 @@ def run_research_process(research_id, query, mode, **kwargs):
from ...security.egress.policy import (
PolicyDeniedError,
context_from_snapshot,
+ resolve_run_primary_engine,
)
try:
- _egress_primary = settings_snapshot.get(
- "search.tool", DEFAULT_SEARCH_TOOL
- )
- if (
- isinstance(_egress_primary, dict)
- and "value" in _egress_primary
- ):
- _egress_primary = _egress_primary["value"]
+ _egress_primary = resolve_run_primary_engine(settings_snapshot)
_egress_ctx = context_from_snapshot(
settings_snapshot,
- _egress_primary or DEFAULT_SEARCH_TOOL,
+ _egress_primary,
username=username,
)
_set_egress_ctx(_egress_ctx)
diff --git a/src/local_deep_research/web_search_engines/search_engine_factory.py b/src/local_deep_research/web_search_engines/search_engine_factory.py
index eb0791d263..cf2c85fe9f 100644
--- a/src/local_deep_research/web_search_engines/search_engine_factory.py
+++ b/src/local_deep_research/web_search_engines/search_engine_factory.py
@@ -3,7 +3,6 @@ from typing import Any, Dict, Optional
from loguru import logger
-from ..utilities.type_utils import unwrap_setting
from ..security.egress.policy import PolicyDeniedError
from ..security.module_whitelist import get_safe_module_class
from .retriever_registry import retriever_registry
@@ -57,6 +56,7 @@ def create_search_engine(
Decision,
context_from_snapshot,
evaluate_retriever,
+ resolve_run_primary_engine,
)
if not settings_snapshot:
@@ -65,13 +65,13 @@ def create_search_engine(
target=engine_name,
)
- primary_raw = unwrap_setting(
- settings_snapshot.get("search.tool", engine_name)
+ primary = resolve_run_primary_engine(
+ settings_snapshot, default=engine_name
)
try:
_ctx = context_from_snapshot(
settings_snapshot,
- primary_raw or engine_name,
+ primary,
username=username,
)
except PolicyDeniedError:
@@ -204,12 +204,15 @@ def create_search_engine(
Decision,
context_from_snapshot,
evaluate_engine,
+ resolve_run_primary_engine,
)
- primary_engine_raw = unwrap_setting(
- settings_snapshot.get("search.tool", engine_name)
+ # Resolve the run's primary via the shared helper so the scope this PEP
+ # enforces matches the one the LangGraph tool-list filter pre-filters with
+ # (default: this engine when search.tool is unset).
+ primary_engine = resolve_run_primary_engine(
+ settings_snapshot, default=engine_name
)
- primary_engine = primary_engine_raw or engine_name
try:
ctx = context_from_snapshot(
diff --git a/tests/security/test_egress_langgraph_tool_filter.py b/tests/security/test_egress_langgraph_tool_filter.py
index 979dc4d7ba..3478ca8cdb 100644
--- a/tests/security/test_egress_langgraph_tool_filter.py
+++ b/tests/security/test_egress_langgraph_tool_filter.py
@@ -87,7 +87,12 @@ def _patched_tool_factories():
def _build_tool_names(
- scope, available, *, retriever_meta=None, snapshot_extra=None
+ scope,
+ available,
+ *,
+ retriever_meta=None,
+ snapshot_extra=None,
+ primary="searxng",
):
"""Construct a minimal strategy under ``scope`` and run the REAL
``_build_tools`` loop against the mocked engine discovery ``available``.
@@ -98,11 +103,20 @@ def _build_tool_names(
``retriever_meta`` (e.g. ``{"is_local": True}``) patches the retriever
registry's metadata lookup so the retriever branch can be exercised.
+
+ ``primary`` sets ``search.tool`` (a real run always has one). Pass
+ ``primary=None`` to omit it entirely and exercise the fail-closed path.
"""
snapshot = {
"policy.egress_scope": scope,
"search.fetch.mode": "disabled",
}
+ # A real run always has a configured primary. Explicit scopes ignore it
+ # (they don't resolve ADAPTIVE); the ADAPTIVE tests override it via
+ # snapshot_extra to drive scope resolution. primary=None omits it so
+ # resolve_run_primary_engine fails closed (raises) — by design.
+ if primary is not None:
+ snapshot["search.tool"] = primary
if snapshot_extra:
snapshot.update(snapshot_extra)
@@ -260,3 +274,77 @@ class TestRetrieverFilter:
assert self._RET in names
assert _specialized(names) == {self._RET}
assert spec.call_args_list[0].args[0] == self._RET
+
+
+# ---------------------------------------------------------------------------
+# ADAPTIVE (the DEFAULT scope): the concrete scope FOLLOWS the run's primary
+# engine (search.tool) — the same value the factory PEP uses. These cover the
+# real-world bug the explicit-scope tests above miss: a private collection
+# primary must pull the run into PRIVATE_ONLY so public specialized engines
+# never reach the agent. Regression for the silent under-filter where
+# _build_egress_context derived the primary from the engine CLASS name instead
+# of search.tool — a collection primary classified as unknown -> BOTH -> public
+# engines (e.g. pubmed) stayed in the agent's tool list, and the factory then
+# hard-denied them mid-run (scope_mismatch_private_only).
+# ---------------------------------------------------------------------------
+
+
+class TestAdaptiveScopeFollowsPrimary:
+ def test_private_primary_excludes_public_engine(self):
+ names, spec = _build_tool_names(
+ "adaptive",
+ {
+ _PUBLIC_ENGINE: {"description": "papers"},
+ _PRIVATE_ENGINE: {"description": "docs"},
+ },
+ snapshot_extra={"search.tool": _PRIVATE_ENGINE},
+ )
+ # ADAPTIVE + private primary => PRIVATE_ONLY: the public engine is
+ # filtered before it ever reaches the agent (name absent AND never
+ # constructed), the local engine survives.
+ assert _PUBLIC_ENGINE not in names
+ called = {c.args[0] for c in spec.call_args_list}
+ assert _PUBLIC_ENGINE not in called
+ assert _specialized(names) == {_PRIVATE_ENGINE}
+
+ def test_public_primary_excludes_private_engine(self):
+ names, spec = _build_tool_names(
+ "adaptive",
+ {
+ _PUBLIC_ENGINE: {"description": "papers"},
+ _PRIVATE_ENGINE: {"description": "docs"},
+ },
+ snapshot_extra={"search.tool": _PUBLIC_ENGINE},
+ )
+ # ADAPTIVE + public primary => PUBLIC_ONLY: the mirror image — the
+ # local engine is filtered, the public one survives.
+ assert _PRIVATE_ENGINE not in names
+ called = {c.args[0] for c in spec.call_args_list}
+ assert _PRIVATE_ENGINE not in called
+ assert _specialized(names) == {_PUBLIC_ENGINE}
+
+
+# ---------------------------------------------------------------------------
+# No configured primary: resolve_run_primary_engine raises ValueError inside
+# _build_egress_context, which catches it and returns None — the advisory
+# tool-list filter degrades to UNFILTERED (the factory PEP still enforces at
+# instantiation) instead of crashing the agent. This locks in the fail-closed
+# contract so the resolve call can't be moved back outside the try/except.
+# ---------------------------------------------------------------------------
+
+
+class TestMissingPrimaryDegradesToUnfiltered:
+ def test_no_primary_does_not_crash_and_leaves_tools_unfiltered(self):
+ # Restrictive scope + NO search.tool. Without the in-try catch this
+ # would raise out of _build_tools; with it, policy_ctx is None so the
+ # filter loop never runs and every discovered engine survives (the
+ # public one is refused later by the factory PEP, not here).
+ names, _ = _build_tool_names(
+ "private_only",
+ {
+ _PUBLIC_ENGINE: {"description": "papers"},
+ _PRIVATE_ENGINE: {"description": "docs"},
+ },
+ primary=None,
+ )
+ assert _specialized(names) == {_PUBLIC_ENGINE, _PRIVATE_ENGINE}
diff --git a/tests/security/test_egress_policy.py b/tests/security/test_egress_policy.py
index b9d02d772d..54e52c4788 100644
--- a/tests/security/test_egress_policy.py
+++ b/tests/security/test_egress_policy.py
@@ -18,6 +18,7 @@ from local_deep_research.security.egress.policy import (
evaluate_llm_endpoint,
evaluate_url,
filter_engines_by_egress,
+ resolve_run_primary_engine,
)
@@ -2474,3 +2475,99 @@ def test_fetch_quota_falls_back_to_local_ctx_without_run_context():
for _ in range(5):
evaluate_url("http://10.0.0.5/", ctx)
assert ctx._fetch_denial_count["count"] == 5
+
+
+# ---------------------------------------------------------------------------
+# resolve_run_primary_engine: the single source of truth for a run's primary
+# engine (what ADAPTIVE classifies). Every run-scoped EgressContext builder
+# routes through this so the scope can't be resolved inconsistently across
+# layers (factory PEP vs. LangGraph tool-list pre-filter).
+# ---------------------------------------------------------------------------
+
+
+class TestResolveRunPrimaryEngine:
+ def test_reads_flat_search_tool(self):
+ assert resolve_run_primary_engine({"search.tool": "pubmed"}) == "pubmed"
+
+ def test_reads_nested_value_shape(self):
+ # The {"value": ...} settings shape is unwrapped like everywhere else.
+ assert (
+ resolve_run_primary_engine({"search.tool": {"value": "paperless"}})
+ == "paperless"
+ )
+
+ # --- Fail closed: a missing/empty primary must NOT silently become a
+ # default (public) engine, which would set the egress scope from an engine
+ # the user never chose. Run-level callers pass no default and must raise.
+
+ def test_missing_key_raises(self):
+ with pytest.raises(ValueError, match="no primary search engine"):
+ resolve_run_primary_engine({})
+
+ def test_none_snapshot_raises(self):
+ with pytest.raises(ValueError, match="no primary search engine"):
+ resolve_run_primary_engine(None)
+
+ def test_empty_value_raises(self):
+ with pytest.raises(ValueError, match="no primary search engine"):
+ resolve_run_primary_engine({"search.tool": ""})
+
+ def test_blank_nested_value_raises(self):
+ # A {"value": ""} shape unwraps to "" — still no configured primary.
+ with pytest.raises(ValueError, match="no primary search engine"):
+ resolve_run_primary_engine({"search.tool": {"value": ""}})
+
+ # --- Explicit default escape hatch: the factory passes default=engine_name
+ # because it is evaluating that one specific engine.
+
+ def test_custom_default_used_when_missing(self):
+ assert resolve_run_primary_engine({}, default="arxiv") == "arxiv"
+
+ def test_custom_default_used_when_empty(self):
+ assert (
+ resolve_run_primary_engine({"search.tool": ""}, default="arxiv")
+ == "arxiv"
+ )
+
+ def test_explicit_value_wins_over_custom_default(self):
+ assert (
+ resolve_run_primary_engine(
+ {"search.tool": "wikipedia"}, default="arxiv"
+ )
+ == "wikipedia"
+ )
+
+ # --- A truthy-but-unusable primary (whitespace / non-string) must NOT
+ # slip past the fail-closed guard and classify to the permissive BOTH.
+
+ def test_whitespace_only_raises(self):
+ with pytest.raises(ValueError, match="no primary search engine"):
+ resolve_run_primary_engine({"search.tool": " "})
+
+ def test_whitespace_is_stripped(self):
+ assert (
+ resolve_run_primary_engine({"search.tool": " pubmed "})
+ == "pubmed"
+ )
+
+ def test_non_string_raises(self):
+ for bad in (5, True, ["arxiv"], {"nested": 1}):
+ with pytest.raises(ValueError, match="no primary search engine"):
+ resolve_run_primary_engine({"search.tool": bad})
+
+ def test_nested_whitespace_value_raises(self):
+ with pytest.raises(ValueError, match="no primary search engine"):
+ resolve_run_primary_engine({"search.tool": {"value": " "}})
+
+ def test_empty_default_is_treated_as_no_default(self):
+ # default="" must not satisfy the fail-closed guard (the old
+ # `if default is not None` would have returned "").
+ with pytest.raises(ValueError, match="no primary search engine"):
+ resolve_run_primary_engine({}, default="")
+
+ def test_whitespace_default_falls_through_to_value(self):
+ # A blank default doesn't mask a real primary.
+ assert (
+ resolve_run_primary_engine({"search.tool": "arxiv"}, default=" ")
+ == "arxiv"
+ )
diff --git a/tests/web/queue/test_processor_v2_real.py b/tests/web/queue/test_processor_v2_real.py
index a9fcc91252..81122d1fb8 100644
--- a/tests/web/queue/test_processor_v2_real.py
+++ b/tests/web/queue/test_processor_v2_real.py
@@ -633,6 +633,42 @@ class TestStartResearch:
assert call_kwargs["model"] == "claude-3"
assert call_kwargs["settings_snapshot"] == {}
+ @patch(f"{MODULE}.start_research_process")
+ def test_legacy_format_seeds_search_tool_from_search_engine(
+ self, mock_start_research
+ ):
+ """A legacy-flat row that carries a search_engine seeds the (otherwise
+ empty) settings_snapshot with search.tool, so the worker's egress build
+ (resolve_run_primary_engine) doesn't fail closed on an empty snapshot
+ and refuse a replayed pre-upgrade queued run."""
+ from local_deep_research.constants import ResearchStatus
+
+ processor = _make_processor()
+ mock_thread = Mock()
+ mock_thread.ident = 1
+ mock_start_research.return_value = mock_thread
+
+ mock_session = MagicMock()
+ mock_research = Mock()
+ mock_research.status = ResearchStatus.QUEUED
+ mock_session.query.return_value.filter_by.return_value.first.return_value = mock_research
+
+ queued = Mock()
+ queued.research_id = "r-102"
+ queued.query = "legacy query"
+ queued.mode = "standard"
+ queued.settings_snapshot = {
+ "model_provider": "anthropic",
+ "search_engine": "pubmed",
+ }
+
+ with patch(f"{MODULE}.UserActiveResearch"):
+ processor._start_research(mock_session, "alice", "pw", queued)
+
+ call_kwargs = mock_start_research.call_args.kwargs
+ assert call_kwargs["settings_snapshot"] == {"search.tool": "pubmed"}
+ assert call_kwargs["search_engine"] == "pubmed"
+
@patch(f"{MODULE}.start_research_process")
def test_none_settings_snapshot_handled(self, mock_start_research):
"""None settings_snapshot treated as empty dict."""
diff --git a/tests/web/routes/test_research_routes_policy.py b/tests/web/routes/test_research_routes_policy.py
index 683e240dbb..806c6e3ddc 100644
--- a/tests/web/routes/test_research_routes_policy.py
+++ b/tests/web/routes/test_research_routes_policy.py
@@ -63,6 +63,11 @@ class TestApplyPolicyOverrides:
def _mgr(snapshot, primary="arxiv"):
m = Mock()
+ # A real settings snapshot always carries the primary under "search.tool".
+ # The precheck now resolves the primary FROM the snapshot (matching the
+ # worker via resolve_run_primary_engine), so reflect that here.
+ if isinstance(snapshot, dict) and "search.tool" not in snapshot:
+ snapshot = {**snapshot, "search.tool": primary}
m.get_settings_snapshot.return_value = snapshot
m.get_setting.side_effect = lambda key, default=None: (
primary if key == "search.tool" else default
@@ -103,6 +108,15 @@ class TestPrecheckEnginePolicy:
mgr = _mgr(None)
assert _precheck_engine_policy(mgr, {}, "library", "user") is None
+ def test_missing_primary_returns_400(self, app_ctx):
+ # No configured primary (empty search.tool) => the precheck fails
+ # CLOSED at the API boundary (400), matching the worker — no silent
+ # searxng fallback that would accept a run the worker then refuses.
+ mgr = _mgr({"policy.egress_scope": "public_only", "search.tool": ""})
+ result = _precheck_engine_policy(mgr, {}, "arxiv", "user")
+ assert result is not None
+ assert result[1] == 400
+
def test_per_research_override_tightens_scope(self, app_ctx):
# Saved scope is permissive (both) but the form overrides to
# public_only for THIS run => a local engine must be refused.
diff --git a/tests/web/services/test_research_service_coverage.py b/tests/web/services/test_research_service_coverage.py
index aa49a6434f..c62d5bdf35 100644
--- a/tests/web/services/test_research_service_coverage.py
+++ b/tests/web/services/test_research_service_coverage.py
@@ -677,7 +677,7 @@ class TestRunResearchProcessQuickMode:
kwargs = dict(
username="alice",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
**extra_kwargs,
)
@@ -967,7 +967,7 @@ class TestRunResearchProcessQuickMode:
"test",
"quick",
username="alice",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
model="m",
search_engine="s",
)
@@ -1082,7 +1082,7 @@ class TestRunResearchProcessDetailedMode:
"query",
"detailed",
username="alice",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
model="m",
search_engine="s",
)
@@ -1212,7 +1212,7 @@ class TestRunResearchProcessDetailedMode:
"query",
"detailed",
username="alice",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
model="m",
search_engine="s",
)
@@ -1302,7 +1302,7 @@ class TestRunResearchProcessDetailedMode:
"query",
"detailed",
username="alice",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
model="m",
search_engine="s",
)
@@ -1370,7 +1370,7 @@ class TestRunResearchProcessSearchErrors:
"query",
"quick",
username="alice",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
model="m",
search_engine="s",
)
@@ -1450,7 +1450,7 @@ class TestRunResearchProcessLLMConfigErrors:
"query",
"quick",
username="alice",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
model="m",
model_provider="llamacpp",
)
@@ -1508,7 +1508,7 @@ class TestRunResearchProcessLLMConfigErrors:
"query",
"quick",
username="alice",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
model="m",
search_engine="searxng",
)
@@ -1572,7 +1572,7 @@ class TestRunResearchProcessErrorHandler:
"query",
"quick",
username="alice",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
model="m",
search_engine="s",
)
@@ -1633,7 +1633,7 @@ class TestRunResearchProcessErrorHandler:
"query",
"quick",
username="alice",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
model="m",
search_engine="s",
)
@@ -1696,7 +1696,7 @@ class TestRunResearchProcessErrorHandler:
"query",
"quick",
username="alice",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
model="m",
search_engine="s",
)
@@ -1759,7 +1759,7 @@ class TestRunResearchProcessErrorHandler:
"query",
"quick",
username="alice",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
model="m",
search_engine="s",
)
@@ -1820,7 +1820,7 @@ class TestRunResearchProcessErrorHandler:
"query",
"quick",
username="alice",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
model="m",
search_engine="s",
)
@@ -1880,7 +1880,7 @@ class TestRunResearchProcessErrorHandler:
"query",
"quick",
username="alice",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
model="m",
search_engine="s",
)
@@ -1953,7 +1953,7 @@ class TestRunResearchProcessResearchContext:
"follow-up query",
"quick",
username="alice",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
model="m",
search_engine="s",
research_context={
@@ -1985,13 +1985,17 @@ class TestRunResearchProcessSettingsContext:
mock_storage.save_report.return_value = True
mock_qp = MagicMock()
- # Settings snapshot with both raw values and setting objects
+ # Settings snapshot with both raw values and setting objects.
+ # Includes search.tool so the run proceeds past the egress build
+ # (which fails closed on a snapshot with no primary) to the
+ # SettingsContext extraction under test.
snapshot = {
"report.citation_format": {
"value": "number_hyperlinks",
"type": "str",
},
"search.max_results": 10, # raw value
+ "search.tool": "searxng",
}
with (
@@ -2132,7 +2136,7 @@ class TestRunResearchProcessSubscription:
"query",
"quick",
username="alice",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
model="m",
search_engine="s",
)
@@ -2210,7 +2214,7 @@ class TestRunResearchProcessFinallyBlock:
"query",
"quick",
username="alice",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
model="m",
search_engine="s",
)
@@ -2288,7 +2292,7 @@ class TestRunResearchProcessFinallyBlock:
"query",
"quick",
username="alice",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
model="m",
search_engine="s",
)
@@ -2361,7 +2365,7 @@ class TestRunResearchProcessSaveReportFailure:
"query",
"quick",
username="alice",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
model="m",
search_engine="s",
)
diff --git a/tests/web/services/test_research_service_coverage_gaps.py b/tests/web/services/test_research_service_coverage_gaps.py
index daef770a24..3d40aceafe 100644
--- a/tests/web/services/test_research_service_coverage_gaps.py
+++ b/tests/web/services/test_research_service_coverage_gaps.py
@@ -245,7 +245,7 @@ class TestSettingsContextInnerClass:
"test",
"quick",
username="user1",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
)
finally:
for p in stack:
@@ -256,6 +256,51 @@ class TestSettingsContextInnerClass:
assert ctx.get_setting("nonexistent.key") is None
+# ---------------------------------------------------------------------------
+# 2b. Worker fails closed when the run has no configured primary engine
+# ---------------------------------------------------------------------------
+
+
+class TestWorkerFailsClosedOnMissingPrimary:
+ """The CLI/scheduler/queue path bypasses the API precheck, so the worker is
+ the only gate. A snapshot with no ``search.tool`` must make
+ run_research_process refuse the run at the egress build
+ (resolve_run_primary_engine raises ValueError) — fail closed — rather than
+ silently running on the public ``searxng`` default. Regression guard for
+ the security-relevant fail-closed change."""
+
+ def test_missing_primary_never_reaches_search_system(self):
+ func = _get_raw_run_research_process()
+
+ mock_system = MagicMock()
+ mock_system.analyze_topic.return_value = {
+ "findings": "x",
+ "formatted_findings": "x",
+ }
+ patches = _base_run_patches()
+ patches[f"{MODULE}.AdvancedSearchSystem"] = MagicMock(
+ return_value=mock_system
+ )
+
+ stack = []
+ for target, mock_obj in patches.items():
+ p = patch(target, mock_obj)
+ stack.append(p)
+ p.start()
+ try:
+ # Empty snapshot => no search.tool => egress build raises =>
+ # run_research_process handles it as a failed run internally.
+ func(1, "test", "quick", username="user1", settings_snapshot={})
+ finally:
+ for p in stack:
+ p.stop()
+
+ # Fail closed: the run aborted before the search system ran. (A
+ # realistic snapshot WITH search.tool does reach analyze_topic — see
+ # TestSettingsContextInnerClass — so this isolates the egress refusal.)
+ mock_system.analyze_topic.assert_not_called()
+
+
# ---------------------------------------------------------------------------
# 3. progress_callback termination check (lines 428-434)
# ---------------------------------------------------------------------------
@@ -303,7 +348,13 @@ class TestProgressCallbackTerminationCheck:
p.start()
try:
# The function catches ResearchTerminatedException internally
- func(1, "test", "quick", username="user1", settings_snapshot={})
+ func(
+ 1,
+ "test",
+ "quick",
+ username="user1",
+ settings_snapshot={"search.tool": "searxng"},
+ )
finally:
for p in stack:
p.stop()
@@ -345,7 +396,7 @@ class TestLLMConfigErrorRaisesValueError:
username="user1",
model=model,
model_provider=provider,
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
)
finally:
for p in stack:
@@ -419,7 +470,7 @@ class TestSearchEngineConfigError:
"quick",
username="user1",
search_engine=search_engine,
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
)
finally:
for p in stack:
@@ -496,7 +547,7 @@ class TestSearchErrorClassification:
"test",
"quick",
username="user1",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
)
finally:
for p in stack:
diff --git a/tests/web/services/test_research_service_deep_coverage.py b/tests/web/services/test_research_service_deep_coverage.py
index ae2c2fdd9a..a304a3a5ea 100644
--- a/tests/web/services/test_research_service_deep_coverage.py
+++ b/tests/web/services/test_research_service_deep_coverage.py
@@ -555,7 +555,11 @@ class TestLLMConfigErrorPaths:
username="user1",
model=model,
model_provider=model_provider,
- settings_snapshot={},
+ # Realistic primary so the run reaches the LLM-error
+ # path under test (not the egress fail-closed raise).
+ settings_snapshot={
+ "search.tool": "searxng"
+ },
)
return mock_qp
diff --git a/tests/web/services/test_research_service_progress_integration.py b/tests/web/services/test_research_service_progress_integration.py
index c495b379ba..a7f0cebc53 100644
--- a/tests/web/services/test_research_service_progress_integration.py
+++ b/tests/web/services/test_research_service_progress_integration.py
@@ -144,6 +144,11 @@ def captured_progress_callback(mode, flask_app):
query="test query",
mode=mode,
username="testuser",
+ # A real run always has a configured primary; without it the
+ # egress context build fails closed (resolve_run_primary_engine
+ # raises) and the worker refuses the run before the callback
+ # wiring under test runs.
+ settings_snapshot={"search.tool": "searxng"},
)
yield captured.get("callback"), progress_state, update_calls