From 58df53c94d05da33a981796fdb46c99b028b501d Mon Sep 17 00:00:00 2001
From: luijait <luijait@icloud.com>
Date: Mon, 29 Sep 2025 17:30:40 +0200
Subject: [PATCH] Agent as a tool fix

---
 .../agents/models/openai_chatcompletions.py   | 27 ++++++-----
 src/cai/util.py                               | 48 ++++++++++++++++++-
 2 files changed, 61 insertions(+), 14 deletions(-)

diff --git a/src/cai/sdk/agents/models/openai_chatcompletions.py b/src/cai/sdk/agents/models/openai_chatcompletions.py
index b7e424f9..80706946 100644
--- a/src/cai/sdk/agents/models/openai_chatcompletions.py
+++ b/src/cai/sdk/agents/models/openai_chatcompletions.py
@@ -2534,21 +2534,22 @@ class OpenAIChatCompletionsModel(Model):
         # start by re-fetching self.is_ollama
         self.is_ollama = os.getenv("OLLAMA") is not None and os.getenv("OLLAMA").lower() == "true"
 
-        # IMPORTANT: Include existing message history for context
+        # IMPORTANT: Build the prompt from history and avoid double-adding the current input.
+        # History already contains the latest user/tool messages for this turn.
         converted_messages = []
-        
-        # First, add all existing messages from history
+
+        # Add existing history first (copy to avoid in-place mutations)
         if self.message_history:
             for msg in self.message_history:
-                msg_copy = msg.copy()  # Use copy to avoid modifying original
-                # Remove any existing cache_control to avoid exceeding the 4-block limit
-                if "cache_control" in msg_copy:
-                    del msg_copy["cache_control"]
+                msg_copy = msg.copy()
+                # Remove any existing cache_control to avoid exceeding provider limits
+                msg_copy.pop("cache_control", None)
                 converted_messages.append(msg_copy)
-        
-        # Then convert and add the new input
-        new_messages = self._converter.items_to_messages(input, model_instance=self)
-        converted_messages.extend(new_messages)
+
+        # Only add converted input if there is no history yet (e.g., very first turn)
+        if not self.message_history:
+            new_messages = self._converter.items_to_messages(input, model_instance=self)
+            converted_messages.extend(new_messages)
 
         if system_instructions:
             # Check if we already have a system message
@@ -3233,7 +3234,7 @@ class OpenAIChatCompletionsModel(Model):
         try:
             if stream:
                 # Standard LiteLLM handling for streaming
-                ret = await litellm.acompletion(**kwargs)
+                # Create a single streaming generator; avoid duplicate calls which leak connections
                 stream_obj = await litellm.acompletion(**kwargs)
 
                 response = Response(
@@ -3287,7 +3288,7 @@ class OpenAIChatCompletionsModel(Model):
                 kwargs["messages"] = messages
                 # Retry once, silently
                 if stream:
-                    ret = await litellm.acompletion(**kwargs)
+                    # Retry with a single streaming generator
                     stream_obj = await litellm.acompletion(**kwargs)
                     response = Response(
                         id=FAKE_RESPONSES_ID,
diff --git a/src/cai/util.py b/src/cai/util.py
index 8385d9a5..fa4c4405 100644
--- a/src/cai/util.py
+++ b/src/cai/util.py
@@ -1377,7 +1377,53 @@ def fix_message_list(messages):  # pylint: disable=R0914,R0915,R0912
             i += 2
         else:
             i += 1
-    return processed_messages
+
+    # Final de-duplication pass to remove accidental duplicates in chat history
+    # This addresses cases where the latest user message or tool pair was added twice upstream.
+    deduped = []
+    for msg in processed_messages:
+        if not deduped:
+            deduped.append(msg)
+            continue
+
+        prev = deduped[-1]
+
+        # Normalize messages for comparison (ignore ephemeral fields ordering)
+        def _norm_tool_calls(m):
+            tcs = m.get("tool_calls") or []
+            norm = []
+            for tc in tcs:
+                if isinstance(tc, dict):
+                    fn = tc.get("function") or {}
+                    norm.append((tc.get("type"), fn.get("name"), fn.get("arguments")))
+            return tuple(norm)
+
+        same_role = prev.get("role") == msg.get("role")
+        # Compare content for text messages
+        same_content = str(prev.get("content")) == str(msg.get("content"))
+        # Compare tool ids for tool messages
+        same_tool_id = (
+            prev.get("role") == "tool"
+            and msg.get("role") == "tool"
+            and str(prev.get("tool_call_id")) == str(msg.get("tool_call_id"))
+            and str(prev.get("content")) == str(msg.get("content"))
+        )
+        # Compare assistant tool calls by name/args (IDs may differ by truncation)
+        same_assistant_tool_calls = (
+            prev.get("role") == "assistant"
+            and msg.get("role") == "assistant"
+            and prev.get("tool_calls")
+            and msg.get("tool_calls")
+            and _norm_tool_calls(prev) == _norm_tool_calls(msg)
+        )
+
+        if (same_role and same_content) or same_tool_id or same_assistant_tool_calls:
+            # Skip exact duplicate
+            continue
+
+        deduped.append(msg)
+
+    return deduped
 
 
 def cli_print_tool_call(tool_name="", args="", output="", prefix="  "):