From c7dc81c0c3e11a81e3784893284a3b52431b49eb Mon Sep 17 00:00:00 2001
From: jenkins <jenkins@bstein.dev>
Date: Sat, 15 Aug 2026 13:35:31 -0300
Subject: [PATCH] hermes: sanitize classifier tool history

---
 services/hermes/scripts/classifier_broker.py | 28 +++++++++++---------
 testing/tests/test_hermes_chat_quality.py    | 25 ++++++++++++++++-
 2 files changed, 40 insertions(+), 13 deletions(-)

diff --git a/services/hermes/scripts/classifier_broker.py b/services/hermes/scripts/classifier_broker.py
index bf1b0152..ba0d64c1 100644
--- a/services/hermes/scripts/classifier_broker.py
+++ b/services/hermes/scripts/classifier_broker.py
@@ -79,18 +79,20 @@ def _compact_message(message: dict[str, Any], limit: int) -> dict[str, Any]:
     result = copy.deepcopy(message)
     if "content" in result:
         result["content"] = _compact_content(result["content"], limit)
-    tool_calls = result.get("tool_calls")
-    if isinstance(tool_calls, list):
-        kept: list[dict[str, Any]] = []
-        for call in tool_calls[-4:]:
-            if not isinstance(call, dict):
-                continue
-            item = copy.deepcopy(call)
-            function = item.get("function")
-            if isinstance(function, dict) and isinstance(function.get("arguments"), str):
-                function["arguments"] = _bounded_text(function["arguments"], 600)
-            kept.append(item)
-        result["tool_calls"] = kept
+    # The local judge only classifies the current boundary. OpenAI-compatible
+    # Ollama validates assistant tool-call JSON and tool/result pairing before
+    # inference, so a bounded or selectively retained transcript can become an
+    # invalid conversation even though its text is sufficient for routing.
+    # Preserve tool evidence as plain user text and remove protocol metadata.
+    had_tool_calls = bool(result.pop("tool_calls", None))
+    result.pop("tool_call_id", None)
+    result.pop("name", None)
+    if result.get("role") == "tool":
+        result["role"] = "user"
+        content = result.get("content")
+        result["content"] = f"[tool evidence]\n{content or ''}"
+    elif had_tool_calls and not result.get("content"):
+        result["content"] = "[assistant requested an external tool]"
     return result
 
 
@@ -153,6 +155,8 @@ def compact_payload(payload: dict[str, Any]) -> dict[str, Any]:
     # The judge never needs tools or binary inputs. Switchyard supplies a
     # response schema separately, and that contract must remain untouched.
     result.pop("tools", None)
+    result.pop("tool_choice", None)
+    result.pop("parallel_tool_calls", None)
     return result
 
 
diff --git a/testing/tests/test_hermes_chat_quality.py b/testing/tests/test_hermes_chat_quality.py
index c4ab80f3..78c918fb 100644
--- a/testing/tests/test_hermes_chat_quality.py
+++ b/testing/tests/test_hermes_chat_quality.py
@@ -1229,10 +1229,24 @@ def test_classifier_broker_bounds_history_without_losing_routing_intent(monkeypa
                 "role": "user",
                 "content": "Build and verify the Cassandra release safely.",
             },
-            {"role": "assistant", "content": "working"},
+            {
+                "role": "assistant",
+                "content": None,
+                "tool_calls": [
+                    {
+                        "id": "call_large",
+                        "type": "function",
+                        "function": {
+                            "name": "large_tool",
+                            "arguments": '{"command":"' + ("x" * 5000) + '"}',
+                        },
+                    }
+                ],
+            },
             {
                 "role": "tool",
                 "content": "unbounded test output " * 10000,
+                "tool_call_id": "call_large",
             },
             {
                 "role": "user",
@@ -1243,6 +1257,8 @@ def test_classifier_broker_bounds_history_without_losing_routing_intent(monkeypa
             },
         ],
         "tools": [{"type": "function", "function": {"name": "large_tool"}}],
+        "tool_choice": "auto",
+        "parallel_tool_calls": True,
         "response_format": {"type": "json_object"},
     }
 
@@ -1252,6 +1268,13 @@ def test_classifier_broker_bounds_history_without_losing_routing_intent(monkeypa
     assert compacted["model"] == payload["model"]
     assert compacted["response_format"] == payload["response_format"]
     assert "tools" not in compacted
+    assert "tool_choice" not in compacted
+    assert "parallel_tool_calls" not in compacted
+    assert all(message.get("role") != "tool" for message in compacted["messages"])
+    assert all("tool_call_id" not in message for message in compacted["messages"])
+    assert all("tool_calls" not in message for message in compacted["messages"])
+    assert "[tool evidence]" in encoded
+    assert "[assistant requested an external tool]" in encoded
     assert "Build and verify the Cassandra release safely." in encoded
     assert "Turn this cat into a cute clown." in encoded
     assert "image attachment available to the selected worker" in encoded