@miller-tech/uap 1.20.13 → 1.20.14

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@miller-tech/uap",
-  "version": "1.20.13",
+  "version": "1.20.14",
   "description": "Autonomous AI agent memory system with CLAUDE.md protocol enforcement",
   "type": "module",
   "main": "dist/index.js",

package/tools/agents/scripts/anthropic_proxy.py

@@ -219,7 +219,7 @@ PROXY_MALFORMED_TOOL_GUARDRAIL = os.environ.get(
     "no",
 }
 PROXY_MALFORMED_TOOL_RETRY_MAX = int(
-    os.environ.get("PROXY_MALFORMED_TOOL_RETRY_MAX", "2")
+    os.environ.get("PROXY_MALFORMED_TOOL_RETRY_MAX", "3")
 )
 PROXY_MALFORMED_TOOL_RETRY_MAX_TOKENS = int(
     os.environ.get("PROXY_MALFORMED_TOOL_RETRY_MAX_TOKENS", "2048")
@@ -3890,6 +3890,40 @@ async def _apply_completion_contract_guardrail(
     return retried
 
 
+def _sanitize_assistant_messages_for_retry(messages: list[dict]) -> list[dict]:
+    """Strip malformed tool-like text from assistant messages to prevent copy-contamination.
+
+    Only sanitizes the last 4 assistant messages to avoid excessive processing.
+    """
+    import re
+
+    # Patterns that indicate malformed tool call text in assistant content
+    _TOOL_LIKE_PATTERNS = re.compile(
+        r"<tool_call>.*?</tool_call>"
+        r"|<function_call>.*?</function_call>"
+        r'|\{"name"\s*:\s*"[^"]+"\s*,\s*"arguments"\s*:'
+        r"|```json\s*\{[^}]*\"name\"\s*:",
+        re.DOTALL,
+    )
+
+    result = list(messages)
+    sanitized_count = 0
+    for i in range(len(result) - 1, -1, -1):
+        if sanitized_count >= 4:
+            break
+        msg = result[i]
+        if msg.get("role") != "assistant":
+            continue
+        content = msg.get("content", "")
+        if isinstance(content, str) and _TOOL_LIKE_PATTERNS.search(content):
+            cleaned = _TOOL_LIKE_PATTERNS.sub("", content).strip()
+            if not cleaned:
+                cleaned = "I will use the appropriate tool."
+            result[i] = {**msg, "content": cleaned}
+            sanitized_count += 1
+    return result
+
+
 def _build_malformed_retry_body(
     openai_body: dict,
     anthropic_body: dict,
@@ -3901,7 +3935,11 @@ def _build_malformed_retry_body(
     retry_body = dict(openai_body)
     retry_body["stream"] = False
     retry_body["tool_choice"] = tool_choice
-    retry_body["temperature"] = PROXY_MALFORMED_TOOL_RETRY_TEMPERATURE
+    # Escalate temperature down on successive retries for more deterministic output
+    if total_attempts > 1 and attempt > 1:
+        retry_body["temperature"] = 0.0
+    else:
+        retry_body["temperature"] = PROXY_MALFORMED_TOOL_RETRY_TEMPERATURE
 
     if tool_choice == "required":
         retry_instruction = (
@@ -3922,7 +3960,10 @@ def _build_malformed_retry_body(
     }
     existing_messages = retry_body.get("messages")
     if isinstance(existing_messages, list) and existing_messages:
-        retry_body["messages"] = [*existing_messages, malformed_retry_instruction]
+        # Strip malformed tool-like text from assistant messages to prevent
+        # the model from copying contaminated patterns on retry
+        sanitized = _sanitize_assistant_messages_for_retry(existing_messages)
+        retry_body["messages"] = [*sanitized, malformed_retry_instruction]
 
     if PROXY_MALFORMED_TOOL_RETRY_MAX_TOKENS > 0:
         current_max = int(

package/tools/agents/tests/test_anthropic_proxy_streaming.py

@@ -3377,6 +3377,66 @@ class TestCycleBreakOptions(unittest.TestCase):
         self.assertEqual(monitor.cycling_tool_names, [])
 
 
+class TestMalformedRetryHardening(unittest.TestCase):
+    """Tests for malformed retry improvements: budget, temp escalation, message sanitization."""
+
+    def test_retry_max_default_is_3(self):
+        """Option 1: default retry budget increased from 2 to 3."""
+        self.assertEqual(proxy.PROXY_MALFORMED_TOOL_RETRY_MAX, 3)
+
+    def test_sanitize_assistant_messages_strips_tool_like_text(self):
+        """Option 3: malformed tool-like text stripped from assistant messages on retry."""
+        messages = [
+            {"role": "system", "content": "You are helpful."},
+            {"role": "user", "content": "Run a command"},
+            {"role": "assistant", "content": 'Here is the result <tool_call>{"name": "Bash", "arguments": {"command": "ls"}}</tool_call>'},
+            {"role": "user", "content": "ok"},
+        ]
+        sanitized = proxy._sanitize_assistant_messages_for_retry(messages)
+        # System and user messages unchanged
+        self.assertEqual(sanitized[0]["content"], "You are helpful.")
+        self.assertEqual(sanitized[1]["content"], "Run a command")
+        self.assertEqual(sanitized[3]["content"], "ok")
+        # Assistant message should have tool_call stripped
+        self.assertNotIn("<tool_call>", sanitized[2]["content"])
+        self.assertNotIn("Bash", sanitized[2]["content"])
+
+    def test_sanitize_preserves_clean_assistant_messages(self):
+        """Clean assistant messages are not modified by sanitization."""
+        messages = [
+            {"role": "assistant", "content": "I will read the file for you."},
+        ]
+        sanitized = proxy._sanitize_assistant_messages_for_retry(messages)
+        self.assertEqual(sanitized[0]["content"], "I will read the file for you.")
+
+    def test_sanitize_replaces_empty_content_with_placeholder(self):
+        """If stripping leaves empty content, a placeholder is used."""
+        messages = [
+            {"role": "assistant", "content": '<tool_call>{"name": "Bash", "arguments": {}}</tool_call>'},
+        ]
+        sanitized = proxy._sanitize_assistant_messages_for_retry(messages)
+        self.assertEqual(sanitized[0]["content"], "I will use the appropriate tool.")
+
+    def test_retry_body_uses_sanitized_messages(self):
+        """Retry body messages are sanitized before adding retry instruction."""
+        openai_body = {
+            "messages": [
+                {"role": "system", "content": "sys"},
+                {"role": "user", "content": "do it"},
+                {"role": "assistant", "content": '<tool_call>{"name":"X","arguments":{}}</tool_call>'},
+            ],
+            "tools": [{"type": "function", "function": {"name": "X", "parameters": {}}}],
+        }
+        anthropic_body = {"tools": [{"name": "X", "input_schema": {"type": "object"}}]}
+        retry = proxy._build_malformed_retry_body(
+            openai_body, anthropic_body, attempt=1, total_attempts=3,
+        )
+        # The assistant message should be sanitized
+        assistant_msgs = [m for m in retry["messages"] if m.get("role") == "assistant"]
+        for m in assistant_msgs:
+            self.assertNotIn("<tool_call>", m.get("content", ""))
+
+
 if __name__ == "__main__":
     unittest.main()