From c0d2ed68f700086f51c9dc2442b5aea4142b94ea Mon Sep 17 00:00:00 2001
From: Wesley Simplicio <wesleysimplicio@live.com>
Date: Sat, 9 May 2026 10:35:08 -0300
Subject: [PATCH] fix(model-metadata): align hy3-preview static context length
 with OpenRouter

Problem:
test_hy3_preview_context_length asserted ctx == 256000 but OpenRouter
live metadata reports 262144 (256 * 1024) for tencent/hy3-preview.
Static DEFAULT_CONTEXT_LENGTHS fallback held 256000, so the assertion
flipped depending on whether the live cache was warm. Issue #22268.

Root cause:
DEFAULT_CONTEXT_LENGTHS["hy3-preview"] = 256000 was authored before
OpenRouter exposed the precise 256K boundary. The static fallback now
diverges from the authoritative live value, producing intermittent
test failures and (more importantly) a 6KB slice of dropped context
when the live cache is unavailable (CI w/o network, offline runs).

Fix:
Bump the static fallback to 262144 to match OpenRouter live metadata.
Comment cites the issue so future readers know why the magic number
isn't a round 256000.

Tests:
- Update test_hy3_preview_context_length to assert 262144.
- Add test_hy3_preview_static_fallback_matches_live which asserts
  against DEFAULT_CONTEXT_LENGTHS directly, so the regression catches
  drift even when the live cache shadows the static table.

Stash-verify confirmed: reverting the static value alone fails the
new static-table test (assert 256000 == 262144).
---
 agent/model_metadata.py                            |  6 ++++--
 tests/hermes_cli/test_tencent_tokenhub_provider.py | 14 +++++++++++++-
 2 files changed, 17 insertions(+), 3 deletions(-)

diff --git a/agent/model_metadata.py b/agent/model_metadata.py
index 4df8a6077791..40fbae33a68b 100644
--- a/agent/model_metadata.py
+++ b/agent/model_metadata.py
@@ -210,8 +210,10 @@ def _strip_provider_prefix(model: str) -> str:
     "grok": 131072,             # catch-all (grok-beta, unknown grok-*)
     # Kimi
     "kimi": 262144,
-    # Tencent — Hy3 Preview (Hunyuan) with 256K context window
-    "hy3-preview": 256000,
+    # Tencent — Hy3 Preview (Hunyuan) with 256K context window.
+    # OpenRouter live metadata reports 262144 (256 * 1024); aligning the
+    # static fallback so cache and fallback agree (issue #22268).
+    "hy3-preview": 262144,
     # Nemotron — NVIDIA's open-weights series (128K context across all sizes)
     "nemotron": 131072,
     # Arcee
diff --git a/tests/hermes_cli/test_tencent_tokenhub_provider.py b/tests/hermes_cli/test_tencent_tokenhub_provider.py
index 62cecaeb0c31..2e5e18810c87 100644
--- a/tests/hermes_cli/test_tencent_tokenhub_provider.py
+++ b/tests/hermes_cli/test_tencent_tokenhub_provider.py
@@ -309,7 +309,19 @@ class TestTencentTokenhubContextLength:
     def test_hy3_preview_context_length(self):
         from agent.model_metadata import get_model_context_length
         ctx = get_model_context_length("hy3-preview")
-        assert ctx == 256000
+        # OpenRouter live API returns 262144 (256 * 1024); static fallback
+        # tracks the same value so tests pass whether resolution comes from
+        # live cache or static table (issue #22268).
+        assert ctx == 262144
+
+    def test_hy3_preview_static_fallback_matches_live(self):
+        """Static DEFAULT_CONTEXT_LENGTHS entry must match OpenRouter live (issue #22268).
+
+        Pins the static value so this stays correct even when the live cache
+        is unavailable in test environments (CI without network).
+        """
+        from agent.model_metadata import DEFAULT_CONTEXT_LENGTHS
+        assert DEFAULT_CONTEXT_LENGTHS["hy3-preview"] == 262144
 
 
 # =============================================================================