diff --git a/strix/config/models.py b/strix/config/models.py
index e8544975..f9fcf106 100644
--- a/strix/config/models.py
+++ b/strix/config/models.py
@@ -733,6 +733,25 @@ def _configure_litellm_default(name: str, value: str) -> None:
     setattr(litellm, name, value)
 
 
+def fallback_model_rejection(fallback: str, primary: str, settings: Settings) -> str | None:
+    """Why ``fallback`` cannot stand in for ``primary`` mid-run, or None if it can.
+
+    Provider credentials, SDK route, and each agent's tool wrappers are all set
+    up once from the primary model, so a fallback that needs a different
+    provider or tool schema would be rejected on every request it serves.
+    """
+    if (
+        _split_model_provider(_normalized_model_name(fallback))[0]
+        != (_split_model_provider(_normalized_model_name(primary))[0])
+    ):
+        return "needs a different provider, whose credentials are not configured"
+    if uses_chat_completions_tool_schema(fallback, settings) != uses_chat_completions_tool_schema(
+        primary, settings
+    ):
+        return "needs a different tool schema than the agents are built with"
+    return None
+
+
 def uses_chat_completions_tool_schema(model_name: str, settings: Settings) -> bool:
     """Return whether the resolved SDK route can only receive JSON function tools."""
     if codex.subscription_model(model_name):
diff --git a/strix/core/runner.py b/strix/core/runner.py
index 1012368f..f2212bdc 100644
--- a/strix/core/runner.py
+++ b/strix/core/runner.py
@@ -22,6 +22,7 @@ from strix.config import load_settings
 from strix.config.models import (
     StrixProvider,
     configure_sdk_model_defaults,
+    fallback_model_rejection,
     uses_chat_completions_tool_schema,
 )
 from strix.config.settings import DEFAULT_MAX_TURNS
@@ -177,14 +178,11 @@ async def run_strix_scan(
     coordinator.set_snapshot_path(agents_path)
     fallback_model = settings.llm.fallback_model
     if fallback_model and (
-        uses_chat_completions_tool_schema(fallback_model, settings) != chat_completions_tools
+        rejection := fallback_model_rejection(fallback_model, resolved_model, settings)
     ):
-        # Agents build their tool set once, for the primary model's schema. A
-        # fallback on the other SDK route would reject every replayed turn, so
-        # it can never recover a denied agent — fail fast instead.
         raise RuntimeError(
-            f"STRIX_LLM_FALLBACK '{fallback_model}' uses a different tool schema than "
-            f"'{resolved_model}'. Pick a fallback from the same provider family."
+            f"STRIX_LLM_FALLBACK '{fallback_model}' {rejection}; it could never serve a "
+            f"turn for '{resolved_model}'. Pick a fallback from the same provider family."
         )
     coordinator.configure_denial_fallback(
         fallback_model,