From d7110889cc78bcd0cee553493ce91659b6d00d2a Mon Sep 17 00:00:00 2001
From: zhangkunyuan <zhangkunyuan@cmhi.chinamobile.com>
Date: Wed, 15 Apr 2026 12:37:42 +0800
Subject: [PATCH] fix(compression): add auxiliary context_length override

---
 cli-config.yaml.example                       |  9 ++++++
 hermes_cli/config.py                          |  1 +
 run_agent.py                                  | 12 +++++++
 .../run_agent/test_compression_feasibility.py | 32 +++++++++++++++++++
 4 files changed, 54 insertions(+)

diff --git a/cli-config.yaml.example b/cli-config.yaml.example
index c9e6645bbadcf..c6ac1409a9ce9 100644
--- a/cli-config.yaml.example
+++ b/cli-config.yaml.example
@@ -366,6 +366,15 @@ compression:
 #   web_extract:
 #     provider: "auto"
 #     model: ""
+#
+#   # Context compression summarization model
+#   compression:
+#     provider: "auto"
+#     model: ""
+#     base_url: ""          # optional custom OpenAI-compatible endpoint
+#     api_key: ""           # optional API key for base_url
+#     context_length: null   # optional explicit context window override (tokens)
+#                            # useful when custom endpoints don't expose /models metadata
 
 # =============================================================================
 # Persistent Memory
diff --git a/hermes_cli/config.py b/hermes_cli/config.py
index 8c46f8bba1313..edeac4ea8e2a2 100644
--- a/hermes_cli/config.py
+++ b/hermes_cli/config.py
@@ -466,6 +466,7 @@ def _ensure_hermes_home_managed(home: Path):
             "model": "",
             "base_url": "",
             "api_key": "",
+            "context_length": None,  # optional explicit override for the compression auxiliary runtime
             "timeout": 120,        # seconds — compression summarises large contexts; increase for local models
         },
         "session_search": {
diff --git a/run_agent.py b/run_agent.py
index 333dda3927fda..72de69140af8d 100644
--- a/run_agent.py
+++ b/run_agent.py
@@ -1748,10 +1748,22 @@ def _check_compression_model_feasibility(self) -> None:
 
             aux_base_url = str(getattr(client, "base_url", ""))
             aux_api_key = str(getattr(client, "api_key", ""))
+            aux_config_context_length = None
+            try:
+                from hermes_cli.config import load_config
+                _cfg = load_config()
+                _aux_cfg = (_cfg.get("auxiliary", {}) or {}).get("compression", {})
+                if isinstance(_aux_cfg, dict):
+                    _raw_aux_ctx = _aux_cfg.get("context_length")
+                    if _raw_aux_ctx is not None:
+                        aux_config_context_length = int(_raw_aux_ctx)
+            except Exception:
+                aux_config_context_length = None
             aux_context = get_model_context_length(
                 aux_model,
                 base_url=aux_base_url,
                 api_key=aux_api_key,
+                config_context_length=aux_config_context_length,
             )
 
             threshold = self.context_compressor.threshold_tokens
diff --git a/tests/run_agent/test_compression_feasibility.py b/tests/run_agent/test_compression_feasibility.py
index 0738b1d438de9..038d800214e94 100644
--- a/tests/run_agent/test_compression_feasibility.py
+++ b/tests/run_agent/test_compression_feasibility.py
@@ -308,3 +308,35 @@ def test_run_conversation_clears_warning_after_replay(mock_get_client, mock_ctx_
         agent._compression_warning = None
 
     assert len(callback_events) == 0
+
+
+@patch("agent.auxiliary_client.get_text_auxiliary_client")
+def test_explicit_auxiliary_compression_context_length_override_is_forwarded(mock_get_client):
+    """auxiliary.compression.context_length should override metadata probing for feasibility checks."""
+    agent = _make_agent(main_context=200_000, threshold_percent=0.50)
+    mock_client = MagicMock()
+    mock_client.base_url = "https://example-proxy/v1"
+    mock_client.api_key = "sk-aux"
+    mock_get_client.return_value = (mock_client, "gpt-5.4")
+
+    messages = []
+    agent._emit_status = lambda msg: messages.append(msg)
+
+    captured = {}
+
+    def _capture(model, **kwargs):
+        captured["model"] = model
+        captured.update(kwargs)
+        return 80_000
+
+    with patch("hermes_cli.config.load_config", return_value={"auxiliary": {"compression": {"context_length": 80_000}}}), \
+         patch("agent.model_metadata.get_model_context_length", side_effect=_capture):
+        agent._check_compression_model_feasibility()
+
+    assert captured["model"] == "gpt-5.4"
+    assert captured["base_url"] == "https://example-proxy/v1"
+    assert captured["api_key"] == "sk-aux"
+    assert captured["config_context_length"] == 80_000
+    assert len(messages) == 1
+    assert "80,000" in messages[0]
+    assert "100,000" in messages[0]