fix(llm): resolve grok/ models to xai/ for LiteLLM metadata

LiteLLM maps xAI models only provider-qualified, so neither "grok/grok-4" nor bare "grok-4" resolves: subscription runs fell back to the generic 200k context window and an 8k output cap instead of Grok's 256k/256k.

subscription.litellm_model_name() now owns the routing-prefix -> LiteLLM name mapping (grok/ -> xai/, chatgpt/ -> bare) and context_budget uses it.

tests/test_context_budget.py

@@ -39,6 +39,17 @@ def test_context_window_chatgpt_prefix_skips_provider_auth(
         context_budget._model_info.cache_clear()
 
 
+def test_context_window_grok_prefix_resolves_to_xai() -> None:
+    # LiteLLM maps xAI models only provider-qualified: neither "grok/grok-4" nor
+    # bare "grok-4" resolves, so the subscription prefix becomes "xai/".
+    context_budget._model_info.cache_clear()
+    try:
+        assert context_budget.context_window("grok/grok-4") == 256_000
+        assert context_budget.output_limit("grok/grok-4") == 256_000
+    finally:
+        context_budget._model_info.cache_clear()
+
+
 def test_context_window_unmapped_uses_fallback(monkeypatch: pytest.MonkeyPatch) -> None:
     context_budget._model_info.cache_clear()
 

tests/test_grok_routing.py

@@ -47,6 +47,17 @@ def test_provider_label_names_the_subscription() -> None:
     assert subscription.provider_label("openai/gpt-5.4") is None
 
 
+def test_litellm_model_name_maps_subscription_prefixes() -> None:
+    # Model metadata (context window, output cap) is keyed "xai/…" for Grok and
+    # bare for ChatGPT; the routing prefixes themselves are never LiteLLM keys.
+    assert subscription.litellm_model_name("grok/grok-4") == "xai/grok-4"
+    assert subscription.litellm_model_name("chatgpt/gpt-5.4") == "gpt-5.4"
+    # Non-subscription models pass through untouched.
+    assert subscription.litellm_model_name("xai/grok-4") == "xai/grok-4"
+    assert subscription.litellm_model_name("openai/gpt-5.4") == "openai/gpt-5.4"
+    assert subscription.litellm_model_name(None) is None
+
+
 def test_run_record_reports_grok_provider(monkeypatch) -> None:  # type: ignore[no-untyped-def]
     settings = mock.MagicMock()
     settings.llm.model = "grok/grok-4"