fix: broken tool call after editing the file before saving

gptme/llm/llm_anthropic.py

@@ -200,9 +200,8 @@ def stream(
 def _handle_tools(message_dicts: Iterable[dict]) -> Generator[dict, None, None]:
     for message in message_dicts:
         # Format tool result as expected by the model
-        if message["role"] == "system" and "call_id" in message:
+        if message["role"] == "user" and "call_id" in message:
             modified_message = dict(message)
-            modified_message["role"] = "user"
             modified_message["content"] = [
                 {
                     "type": "tool_result",
@@ -358,22 +357,35 @@ def _transform_system_messages(
     # unless a `call_id` is present, indicating the tool_format is 'tool'.
     # Tool responses are handled separately by _handle_tool.
     for i, message in enumerate(messages):
-        if message.role == "system" and message.call_id is None:
+        if message.role == "system":
+            content = (
+                f"<system>{message.content}</system>"
+                if message.call_id is None
+                else message.content
+            )
+
             messages[i] = Message(
                 "user",
-                content=f"<system>{message.content}</system>",
+                content=content,
                 files=message.files,  # type: ignore
+                call_id=message.call_id,
             )
 
     # find consecutive user role messages and merge them together
     messages_new: list[Message] = []
     while messages:
         message = messages.pop(0)
-        if messages_new and messages_new[-1].role == "user" and message.role == "user":
+        if (
+            messages_new
+            and messages_new[-1].role == "user"
+            and message.role == "user"
+            and message.call_id == messages_new[-1].call_id
+        ):
             messages_new[-1] = Message(
                 "user",
                 content=f"{messages_new[-1].content}\n\n{message.content}",
                 files=messages_new[-1].files + message.files,  # type: ignore
+                call_id=messages_new[-1].call_id,
             )
         else:
             messages_new.append(message)

gptme/llm/llm_openai.py

@@ -274,6 +274,7 @@ def _handle_tools(message_dicts: Iterable[dict]) -> Generator[dict, None, None]:
                 modified_message["content"] = content
 
             if tool_calls:
+                # Clean content property if empty otherwise the call fails
                 if not content:
                     del modified_message["content"]
                 modified_message["tool_calls"] = tool_calls
@@ -283,6 +284,41 @@ def _handle_tools(message_dicts: Iterable[dict]) -> Generator[dict, None, None]:
             yield message
 
 
+def _merge_tool_results_with_same_call_id(
+    messages_dicts: Iterable[dict],
+) -> list[dict]:  # Generator[dict, None, None]:
+    """
+    When we call a tool, this tool can potentially yield multiple messages. However
+    the API expect to have only one tool result per tool call. This function tries
+    to merge subsequent tool results with the same call ID as expected by
+    the API.
+    """
+
+    messages_dicts = iter(messages_dicts)
+
+    messages_new: list[dict] = []
+    while message := next(messages_dicts, None):
+        if messages_new and (
+            message["role"] == "tool"
+            and messages_new[-1]["role"] == "tool"
+            and message["tool_call_id"] == messages_new[-1]["tool_call_id"]
+        ):
+            prev_msg = messages_new[-1]
+            content = message["content"]
+            if not isinstance(content, list):
+                content = {"type": "text", "text": content}
+
+            messages_new[-1] = {
+                "role": "tool",
+                "content": prev_msg["content"] + content,
+                "tool_call_id": prev_msg["tool_call_id"],
+            }
+        else:
+            messages_new.append(message)
+
+    return messages_new
+
+
 def _process_file(msg: dict, model: ModelMeta) -> dict:
     message_content = msg["content"]
     if model.provider == "deepseek":
@@ -423,7 +459,9 @@ def _prepare_messages_for_api(
     tools_dict = [_spec2tool(tool, model) for tool in tools] if tools else None
 
     if tools_dict is not None:
-        messages_dicts = _handle_tools(messages_dicts)
+        messages_dicts = _merge_tool_results_with_same_call_id(
+            _handle_tools(messages_dicts)
+        )
 
     messages_dicts = _transform_msgs_for_special_provider(messages_dicts, model)
 

gptme/tools/__init__.py

@@ -2,6 +2,8 @@
 from collections.abc import Generator
 from functools import lru_cache
 
+from ..util.interrupt import clear_interruptible
+
 from ..message import Message
 from .base import (
     ConfirmFunc,
@@ -124,11 +126,13 @@ def execute_msg(msg: Message, confirm: ConfirmFunc) -> Generator[Message, None,
                 for tool_response in tooluse.execute(confirm):
                     yield tool_response.replace(call_id=tooluse.call_id)
             except KeyboardInterrupt:
+                clear_interruptible()
                 yield Message(
                     "system",
                     "User hit Ctrl-c to interrupt the process",
                     call_id=tooluse.call_id,
                 )
+                break
 
 
 # Called often when checking streaming output for executable blocks,

tests/test_llm_anthropic.py

@@ -97,6 +97,7 @@ def test_message_conversion_with_tools():
             content='<thinking>\nSomething\n</thinking>\n@save(tool_call_id): {"path": "path.txt", "content": "file_content"}',
         ),
         Message(role="system", content="Saved to toto.txt", call_id="tool_call_id"),
+        Message(role="system", content="(Modified by user)", call_id="tool_call_id"),
     ]
 
     tool_save = get_tool("save")
@@ -152,7 +153,12 @@ def test_message_conversion_with_tools():
             "content": [
                 {
                     "type": "tool_result",
-                    "content": [{"type": "text", "text": "Saved to toto.txt"}],
+                    "content": [
+                        {
+                            "type": "text",
+                            "text": "Saved to toto.txt\n\n(Modified by user)",
+                        }
+                    ],
                     "tool_use_id": "tool_call_id",
                     "cache_control": {"type": "ephemeral"},
                 }

tests/test_llm_openai.py

@@ -116,6 +116,7 @@ def test_message_conversion_with_tools():
             content='\n@save(tool_call_id): {"path": "path.txt", "content": "file_content"}',
         ),
         Message(role="system", content="Saved to toto.txt", call_id="tool_call_id"),
+        Message(role="system", content="(Modified by user)", call_id="tool_call_id"),
     ]
 
     set_default_model("openai/gpt-4o")
@@ -193,7 +194,10 @@ def test_message_conversion_with_tools():
         },
         {
             "role": "tool",
-            "content": [{"type": "text", "text": "Saved to toto.txt"}],
+            "content": [
+                {"type": "text", "text": "Saved to toto.txt"},
+                {"type": "text", "text": "(Modified by user)"},
+            ],
             "tool_call_id": "tool_call_id",
         },
     ]