diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 757467a00..b4f8fe1ca 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -67,7 +67,6 @@ from core.tool_healing import (
     _strip_bracket_tag_calls,
     apply_tool_strip_patterns,
     strip_outside_think,
-    strip_tool_call_markup,
 )
 from utils.native_path_leases import child_env_without_native_path_secret
 from utils.hf_xet_fallback import hf_hub_download_with_xet_fallback
@@ -1781,6 +1780,13 @@ class LlamaCppBackend:
             return False
         return self._supports_tools
 
+    @property
+    def supports_tool_passthrough(self) -> bool:
+        # supports_tools is forced off for DiffusionGemma (its agentic loop drops the
+        # per-step canvas frames), but client passthrough skips that loop, so it uses
+        # the real _supports_tools.
+        return self._supports_tools
+
     @property
     def cache_type_kv(self) -> Optional[str]:
         return self._cache_type_kv
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 5332037e0..ce755ae1f 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -6053,14 +6053,13 @@ async def openai_chat_completions(
     # free-form sampling. Guided decoding does not require ``supports_tools`` --
     # the grammar machinery is independent of tool-call parsing.
     _has_response_format = _extract_response_format(payload) is not None
-    _tools_passthrough = llama_backend.supports_tools and (
-        (payload.tools and len(payload.tools) > 0) or _has_tool_messages
-    )
-    if (
-        using_gguf
-        and not _effective_enable_tools(payload)
-        and (_tools_passthrough or _has_response_format)
-    ):
+    _tools_passthrough = getattr(
+        llama_backend, "supports_tool_passthrough", llama_backend.supports_tools
+    ) and ((payload.tools and len(payload.tools) > 0) or _has_tool_messages)
+    # DiffusionGemma keeps supports_tools off, so the server-side tool loop can't
+    # claim the request; fall through to client passthrough, matching /v1/messages.
+    _server_tool_loop = _effective_enable_tools(payload) and llama_backend.supports_tools
+    if using_gguf and not _server_tool_loop and (_tools_passthrough or _has_response_format):
         if _wants_multiple_choices(payload):
             raise _reject_unsupported_n("GGUF tool or response_format passthrough")
         if payload.audio_base64:
@@ -10222,7 +10221,9 @@ async def anthropic_messages(
         and not _has_image
     )
     client_tools = (
-        not server_tools and len(openai_client_tools) > 0 and llama_backend.supports_tools
+        not server_tools
+        and len(openai_client_tools) > 0
+        and getattr(llama_backend, "supports_tool_passthrough", llama_backend.supports_tools)
     )
 
     # Anthropic tool_choice.disable_parallel_tool_use caps the response to a