From 6973aee4df39bd316c4a5eb0aee706314e090bee Mon Sep 17 00:00:00 2001
From: "lone.yao" <cherished-liuxu@users.noreply.github.com>
Date: Fri, 15 May 2026 11:45:26 +0800
Subject: [PATCH] fix: honor custom provider context for auxiliary compression

---
 run_agent.py                                  |  7 ++-
 ...aux_compression_custom_provider_context.py | 61 +++++++++++++++++++
 2 files changed, 67 insertions(+), 1 deletion(-)
 create mode 100644 tests/test_aux_compression_custom_provider_context.py

diff --git a/run_agent.py b/run_agent.py
index 325e1e13ef3c4..e09a9c065f2b9 100644
--- a/run_agent.py
+++ b/run_agent.py
@@ -2260,6 +2260,10 @@ def __init__(
         # Persist for reuse on switch_model / fallback activation. Must come
         # AFTER the custom_providers branch so per-model overrides aren't lost.
         self._config_context_length = _config_context_length
+        # Also persist the compatible custom-provider list so other runtime
+        # checks (notably auxiliary compression feasibility) can reuse the
+        # exact same per-model context_length overrides resolved at startup.
+        self._custom_providers = _custom_providers
 
         self._ensure_lmstudio_runtime_loaded(_config_context_length)
 
@@ -3250,9 +3254,10 @@ def _check_compression_model_feasibility(self) -> None:
                 # provider-specific paths (e.g. Bedrock static table, OpenRouter API)
                 # are invoked for the correct client, not inherited from the main model.
                 provider=(_aux_cfg_provider if _aux_cfg_provider and _aux_cfg_provider != "auto" else getattr(self, "provider", "")),
-                custom_providers=self._custom_providers,
+                custom_providers=getattr(self, "_custom_providers", None),
             )
 
+
             # Hard floor: the auxiliary compression model must have at least
             # MINIMUM_CONTEXT_LENGTH (64K) tokens of context.  The main model
             # is already required to meet this floor (checked earlier in
diff --git a/tests/test_aux_compression_custom_provider_context.py b/tests/test_aux_compression_custom_provider_context.py
new file mode 100644
index 0000000000000..32a98321b48d7
--- /dev/null
+++ b/tests/test_aux_compression_custom_provider_context.py
@@ -0,0 +1,61 @@
+"""Regression test for auxiliary compression context lookup on custom endpoints.
+
+The startup feasibility check used to call ``get_model_context_length()`` for the
+auxiliary compression model without threading through ``custom_providers``.
+That caused named custom endpoints to miss per-model overrides and fall back to
+``DEFAULT_FALLBACK_CONTEXT`` (256K), even when startup had already resolved a
+larger explicit context window for the same endpoint/model.
+"""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+from unittest.mock import patch
+
+from run_agent import AIAgent
+
+
+class _DummyCompressor:
+    def __init__(self, threshold_tokens: int = 500_000, context_length: int = 1_000_000):
+        self.threshold_tokens = threshold_tokens
+        self.threshold_percent = threshold_tokens / context_length
+        self.context_length = context_length
+
+
+def test_aux_compression_feasibility_threads_custom_provider_overrides():
+    agent = object.__new__(AIAgent)
+    agent.compression_enabled = True
+    agent.context_compressor = _DummyCompressor()
+    agent.provider = "gpt"
+    agent.model = "gpt-5.4"
+    agent._aux_compression_context_length_config = None
+    agent._custom_providers = [
+        {
+            "name": "gpt",
+            "base_url": "https://lovethea.org",
+            "models": {
+                "gpt-5.4": {"context_length": 1_000_000},
+            },
+        }
+    ]
+    agent._compression_warning = None
+    agent.status_callback = None
+    agent._emit_status = lambda msg: None
+    agent._current_main_runtime = lambda: {}
+
+    captured = {}
+
+    def _fake_get_model_context_length(*args, **kwargs):
+        captured["kwargs"] = kwargs
+        return 1_000_000
+
+    fake_client = SimpleNamespace(base_url="https://lovethea.org", api_key="redacted")
+
+    with patch("agent.auxiliary_client.get_text_auxiliary_client", return_value=(fake_client, "gpt-5.4")), \
+         patch("agent.auxiliary_client._resolve_task_provider_model", return_value=("gpt", "gpt-5.4", "", "", {})), \
+         patch("agent.model_metadata.get_model_context_length", side_effect=_fake_get_model_context_length):
+        agent._check_compression_model_feasibility()
+
+    assert captured["kwargs"]["custom_providers"] == agent._custom_providers
+    assert agent.context_compressor.threshold_tokens == 500_000
+    assert agent._compression_warning is None