From 53ad7794e5810fa48cb532edf0308c647fb8d0da Mon Sep 17 00:00:00 2001
From: Julientalbot <julien.talbot@ergonomia.re>
Date: Thu, 13 Aug 2026 11:27:44 +0400
Subject: [PATCH] fix(xai): drop stale 256K grok-4.6 context cache

docs.x.ai (2026-08-12): grok-4.6 is the flagship, 500K context.
Live GET /v1/models lists grok-4.6 at context_length 500000
(no grok-4.6-latest alias).

#84661 landed the catalog. Main already lists native grok-4.6
on the xAI picker. This is only the leftover cache guard
(same pattern as grok-4.3): pre-catalog builds persisted the
grok-4 catch-all (256K).
---
 agent/model_metadata.py            | 22 ++++++++++++++++++++++
 tests/agent/test_model_metadata.py | 30 ++++++++++++++++++++++++++++++
 2 files changed, 52 insertions(+)

diff --git a/agent/model_metadata.py b/agent/model_metadata.py
index 8fabeb716d..ca53545f33 100644
--- a/agent/model_metadata.py
+++ b/agent/model_metadata.py
@@ -2021,6 +2021,19 @@ def _model_name_suggests_grok_4_3(model: str) -> bool:
     return "grok-4.3" in model.lower()
 
 
+def _model_name_suggests_grok_4_6(model: str) -> bool:
+    """Return True if the model name looks like a Grok 4.6 variant.
+
+    Catches ``grok-4.6`` and aggregator-prefixed slugs. Used as a guard
+    against stale cache entries seeded by pre-catalog builds that resolved
+    grok-4.6 via the generic ``grok-4`` catch-all (256,000) before the
+    500K entry existed. Official card: docs.x.ai/developers/models/grok-4.6.
+    Live GET /v1/models lists ``grok-4.6`` at context_length 500000 and does
+    not publish a ``grok-4.6-latest`` alias.
+    """
+    return "grok-4.6" in model.lower()
+
+
 def _query_local_context_length(model: str, base_url: str, api_key: str = "") -> Optional[int]:
     """Query a local server for the model's context length (short-TTL cached).
 
@@ -2645,6 +2658,15 @@ def get_model_context_length(
                     model, base_url, f"{cached:,}",
                 )
                 _invalidate_cached_context_length(model, base_url)
+            # Same class of leftover as grok-4.3: grok-4.6 is 500K, but
+            # pre-catalog builds persisted the grok-4 catch-all (256,000).
+            elif cached <= 256_000 and _model_name_suggests_grok_4_6(model):
+                logger.info(
+                    "Dropping stale Grok-4.6 cache entry %s@%s -> %s (pre-catalog value); "
+                    "re-resolving via hardcoded defaults",
+                    model, base_url, f"{cached:,}",
+                )
+                _invalidate_cached_context_length(model, base_url)
             # Nous Portal: the portal /v1/models endpoint is authoritative.
             # Bypass the persistent cache so step 5b can always reconcile
             # against it — this corrects pre-fix entries seeded from the
diff --git a/tests/agent/test_model_metadata.py b/tests/agent/test_model_metadata.py
index bfcce7e67c..9d7ec36ff1 100644
--- a/tests/agent/test_model_metadata.py
+++ b/tests/agent/test_model_metadata.py
@@ -1221,6 +1221,36 @@ class TestGrok43StaleCacheGuard:
             assert ctx == 256_000, f"{slug} should stay 256000, got {ctx}"
 
 
+class TestGrok46StaleCacheGuard:
+    """Pre-catalog builds resolved grok-4.6 via the generic 'grok-4' catch-all
+    (256,000) and persisted it before the 500K catalog entry existed.
+    The step-1 cache guard must drop that stale value and re-resolve to 500K.
+    Official card: 500,000 context (docs.x.ai/developers/models/grok-4.6).
+    """
+
+    def test_suggests_grok_4_6(self):
+        from agent.model_metadata import _model_name_suggests_grok_4_6
+        assert _model_name_suggests_grok_4_6("grok-4.6")
+        assert _model_name_suggests_grok_4_6("xai/grok-4.6")
+        assert _model_name_suggests_grok_4_6("x-ai/grok-4.6")
+        assert not _model_name_suggests_grok_4_6("grok-4")
+        assert not _model_name_suggests_grok_4_6("grok-4-fast")
+        assert not _model_name_suggests_grok_4_6("grok-4.5")
+        assert not _model_name_suggests_grok_4_6("grok-4.3")
+
+    def test_stale_grok_4_6_dropped_and_reresolves_to_500k(self, tmp_path, monkeypatch):
+        monkeypatch.setenv("HERMES_HOME", str(tmp_path))
+        import importlib
+        import agent.model_metadata as mm
+        importlib.reload(mm)
+        base = "https://api.x.ai/v1"
+        mm.save_context_length("grok-4.6", base, 256_000)
+        ctx = mm.get_model_context_length(
+            "grok-4.6", base_url=base, api_key="", provider="xai"
+        )
+        assert ctx == 500_000
+
+
 class TestMoAContextLength:
     """MoA virtual provider resolves context from the aggregator slot, not 256K default."""