From 585899f1cc7def04792f6291916f706bf81db74f Mon Sep 17 00:00:00 2001
From: Yucheng He <yucheng@berri.ai>
Date: Sat, 5 Sep 2026 14:21:45 -0700
Subject: [PATCH] feat: default capabilities for new chat catalog entries

---
 scripts/sync_together_ai_models.py                 | 11 +++++++++++
 tests/test_litellm/test_sync_together_ai_models.py |  9 ++++++++-
 2 files changed, 19 insertions(+), 1 deletion(-)

diff --git a/scripts/sync_together_ai_models.py b/scripts/sync_together_ai_models.py
index 12b128890f1..e77cb475cd5 100644
--- a/scripts/sync_together_ai_models.py
+++ b/scripts/sync_together_ai_models.py
@@ -85,6 +85,16 @@ _TOOLS: Final = MappingProxyType(
     }
 )
 
+_CHAT_DEFAULTS: Final = MappingProxyType(
+    {
+        "max_output_tokens": 120000,
+        "supports_function_calling": True,
+        "supports_parallel_function_calling": True,
+        "supports_reasoning": True,
+        "supports_tool_choice": True,
+    }
+)
+
 CAPABILITY_RULES: Final = (
     _rule(
         "MiniMaxAI/MiniMax-M3",
@@ -308,6 +318,7 @@ def _new_entry(model: CatalogModel, mode: str) -> RegistryEntry:
         "litellm_provider": PROVIDER,
         "mode": mode,
         "source": SOURCE_URL,
+        **(_CHAT_DEFAULTS if model.type == "chat" else {}),
         **(dict(rule.fields) if rule else {}),
     }
     return dict(sorted(merged.items()))
diff --git a/tests/test_litellm/test_sync_together_ai_models.py b/tests/test_litellm/test_sync_together_ai_models.py
index b8a85bcfbdc..f8ee2471da4 100644
--- a/tests/test_litellm/test_sync_together_ai_models.py
+++ b/tests/test_litellm/test_sync_together_ai_models.py
@@ -105,6 +105,7 @@ def test_added_chat_model_matches_reviewed_registry_shape() -> None:
         "input_cost_per_token": 3e-06,
         "litellm_provider": "together_ai",
         "max_input_tokens": 1048576,
+        "max_output_tokens": 120000,
         "max_tokens": 1048576,
         "mode": "chat",
         "output_cost_per_token": 1.5e-05,
@@ -119,6 +120,12 @@ def test_added_chat_model_matches_reviewed_registry_shape() -> None:
     }
 
 
+def test_unreviewed_chat_model_gets_catalog_defaults() -> None:
+    outcome = sync.compute_sync({}, [_chat_model("acme/unreviewed")], _doc({"x": "2026-01-01"}))
+    entry = outcome.cost_map["together_ai/acme/unreviewed"]
+    assert {key: entry[key] for key in sync._CHAT_DEFAULTS} == dict(sync._CHAT_DEFAULTS)
+
+
 def test_added_embedding_model_has_no_output_token_cap() -> None:
     outcome = sync.compute_sync({}, RECORDED_CATALOG, RECORDED_DOC)
     assert outcome.cost_map["together_ai/intfloat/multilingual-e5-large-instruct"] == {
@@ -144,7 +151,7 @@ def test_output_ceiling_comes_from_the_rule_never_from_context_length() -> None:
     glm = next(model for model in RECORDED_CATALOG if model.id == "zai-org/GLM-5.2")
     fresh = sync.compute_sync({}, [_chat_model("acme/unreviewed", ctx=1048576), glm], _doc({"x": "2026-01-01"}))
     unreviewed = fresh.cost_map["together_ai/acme/unreviewed"]
-    assert "max_output_tokens" not in unreviewed
+    assert unreviewed["max_output_tokens"] == 120000
     assert (unreviewed["max_input_tokens"], unreviewed["max_tokens"]) == (1048576, 1048576)
     reviewed = fresh.cost_map["together_ai/zai-org/GLM-5.2"]
     assert (reviewed["max_input_tokens"], reviewed["max_output_tokens"], reviewed["max_tokens"]) == (