feat: use the customtools model for gemini

core/framework/agent_loop/internals/tool_result_handler.py

@@ -217,7 +217,7 @@ def truncate_tool_result(
     When *spillover_dir* is configured, EVERY non-error tool result is
     written to disk for debugging. The LLM-visible content is then
     shaped to avoid a **poison pattern** that we traced on 2026-04-15
-    through a gemini-3.1-pro-preview queen session: the prior format
+    through a gemini-3.1-pro-preview-customtools queen session: the prior format
     appended ``\\n\\n[Saved to '/abs/path/file.txt']`` after every
     small result, and frontier pattern-matching models (gemini 3.x in
     particular) learned to autocomplete the `[Saved to '...']` trailer

core/framework/llm/model_catalog.json

@@ -64,7 +64,7 @@
           "max_context_tokens": 900000
         },
         {
-          "id": "gemini-3.1-pro-preview",
+          "id": "gemini-3.1-pro-preview-customtools",
           "label": "Gemini 3.1 Pro - Best quality",
           "recommended": true,
           "max_tokens": 32768,
@@ -305,7 +305,7 @@
           "max_context_tokens": 872000
         },
         {
-          "id": "google/gemini-3.1-pro-preview",
+          "id": "google/gemini-3.1-pro-preview-customtools",
           "label": "Gemini 3.1 Pro Preview - Long-context reasoning",
           "recommended": false,
           "max_tokens": 32768,

core/tests/test_model_catalog.py

@@ -150,7 +150,7 @@ def test_openrouter_catalog_tracks_current_frontier_set():
         "openai/gpt-5.4",
         "anthropic/claude-sonnet-4.6",
         "anthropic/claude-opus-4.6",
-        "google/gemini-3.1-pro-preview",
+        "google/gemini-3.1-pro-preview-customtools",
         "deepseek/deepseek-v3.2",
     ]
     assert openrouter_models[0]["max_tokens"] == 128000