fix: correctly indent agent tool execution block inside the round loop to prevent duplicate calls and rate limits

backend.py

@@ -836,92 +836,92 @@ async def chat_completions(request: Request, authorization: str = Header(None)):
                         if choice.finish_reason == "tool_calls":
                             break
 
-                # If no tool calls, we're done
-                if not tool_calls_raw:
-                    await save_message(session_id, "assistant", full_content)
-                    yield make_chunk(request_id, requested_model, finish_reason="stop")
-                    yield "data: [DONE]\n\n"
-                    return
-
-                # Execute tool calls
-                tool_calls_list = []
-                for idx in sorted(tool_calls_raw.keys()):
-                    tc = tool_calls_raw[idx]
-                    tool_calls_list.append({
-                        "id": tc["id"],
-                        "type": "function",
-                        "function": {"name": tc["name"], "arguments": tc["arguments"]}
-                    })
-
-                # Add assistant message with tool calls to history
-                assistant_msg = {"role": "assistant", "content": full_content or None, "tool_calls": tool_calls_list}
-                final_messages.append(assistant_msg)
-
-                # Execute each tool and add results
-                for tc in tool_calls_list:
-                    func_name = tc["function"]["name"]
-                    raw_args_str = tc["function"]["arguments"]
-                    try:
-                        func_args = json.loads(raw_args_str)
-                    except json.JSONDecodeError:
-                        # Attempt raw JSON repair
-                        repaired_str = raw_args_str.strip()
-                        if not repaired_str.startswith("{"):
-                            repaired_str = "{" + repaired_str
-                        if not repaired_str.endswith("}"):
-                            repaired_str = repaired_str + "}"
+                    # If no tool calls, we're done
+                    if not tool_calls_raw:
+                        await save_message(session_id, "assistant", full_content)
+                        yield make_chunk(request_id, requested_model, finish_reason="stop")
+                        yield "data: [DONE]\n\n"
+                        return
+
+                    # Execute tool calls
+                    tool_calls_list = []
+                    for idx in sorted(tool_calls_raw.keys()):
+                        tc = tool_calls_raw[idx]
+                        tool_calls_list.append({
+                            "id": tc["id"],
+                            "type": "function",
+                            "function": {"name": tc["name"], "arguments": tc["arguments"]}
+                        })
+
+                    # Add assistant message with tool calls to history
+                    assistant_msg = {"role": "assistant", "content": full_content or None, "tool_calls": tool_calls_list}
+                    final_messages.append(assistant_msg)
+
+                    # Execute each tool and add results
+                    for tc in tool_calls_list:
+                        func_name = tc["function"]["name"]
+                        raw_args_str = tc["function"]["arguments"]
                         try:
-                            func_args = json.loads(repaired_str)
-                            log_activity(f"Auto-fixed invalid JSON string for tool: {func_name}")
-                        except:
-                            func_args = {}
+                            func_args = json.loads(raw_args_str)
+                        except json.JSONDecodeError:
+                            # Attempt raw JSON repair
+                            repaired_str = raw_args_str.strip()
+                            if not repaired_str.startswith("{"):
+                                repaired_str = "{" + repaired_str
+                            if not repaired_str.endswith("}"):
+                                repaired_str = repaired_str + "}"
+                            try:
+                                func_args = json.loads(repaired_str)
+                                log_activity(f"Auto-fixed invalid JSON string for tool: {func_name}")
+                            except:
+                                func_args = {}
 
-                    # Perform semantic repairs
-                    repaired_args, repair_notes = repair_arguments(func_name, func_args)
+                        # Perform semantic repairs
+                        repaired_args, repair_notes = repair_arguments(func_name, func_args)
                     
-                    # Log activity
-                    log_activity(f"Tool execution: {func_name} args={repaired_args}")
-                    if repair_notes:
-                        for note in repair_notes:
-                            log_activity(f"[Tool Repair] {note}")
-
-                    # Show tool execution to user
-                    yield make_chunk(request_id, requested_model, f"\n\n🔧 **{func_name}**")
-                    if repair_notes:
-                        yield make_chunk(request_id, requested_model, " *(Auto-Repaired)*")
+                        # Log activity
+                        log_activity(f"Tool execution: {func_name} args={repaired_args}")
+                        if repair_notes:
+                            for note in repair_notes:
+                                log_activity(f"[Tool Repair] {note}")
+
+                        # Show tool execution to user
+                        yield make_chunk(request_id, requested_model, f"\n\n🔧 **{func_name}**")
+                        if repair_notes:
+                            yield make_chunk(request_id, requested_model, " *(Auto-Repaired)*")
                     
-                    if func_name == "run_bash" and "command" in repaired_args:
-                        yield make_chunk(request_id, requested_model, f": `{repaired_args['command']}`\n")
-                    elif func_name == "read_file" and "path" in repaired_args:
-                        yield make_chunk(request_id, requested_model, f": `{repaired_args['path']}`\n")
-                    elif func_name == "write_file" and "path" in repaired_args:
-                        yield make_chunk(request_id, requested_model, f": `{repaired_args['path']}`\n")
-                    elif func_name == "list_directory":
-                        yield make_chunk(request_id, requested_model, f": `{repaired_args.get('path', '.')}`\n")
-                    elif func_name == "grep_search":
-                        yield make_chunk(request_id, requested_model, f": `{repaired_args.get('pattern', '')}`\n")
-                    else:
-                        yield make_chunk(request_id, requested_model, "\n")
-
-                    # Execute the tool
-                    result = await execute_tool(func_name, repaired_args)
-
-                    # Append teaching note if repaired
-                    if repair_notes:
-                        result += f"\n\n[SYSTEM REPAIR NOTE: The harness automatically fixed formatting issues: {', '.join(repair_notes)}. Please strictly follow the tool's JSON schema in subsequent calls without these wrapping/formatting errors.]"
-
-                    # Show truncated result to user
-                    preview = result[:500] + ("..." if len(result) > 500 else "")
-                    yield make_chunk(request_id, requested_model, f"```\n{preview}\n```\n")
-
-                    # Add tool result to message history
-                    final_messages.append({
-                        "role": "tool",
-                        "tool_call_id": tc["id"],
-                        "content": result,
-                    })
-
-                    await save_message(session_id, "tool", result, tool_call_id=tc["id"])
+                        if func_name == "run_bash" and "command" in repaired_args:
+                            yield make_chunk(request_id, requested_model, f": `{repaired_args['command']}`\n")
+                        elif func_name == "read_file" and "path" in repaired_args:
+                            yield make_chunk(request_id, requested_model, f": `{repaired_args['path']}`\n")
+                        elif func_name == "write_file" and "path" in repaired_args:
+                            yield make_chunk(request_id, requested_model, f": `{repaired_args['path']}`\n")
+                        elif func_name == "list_directory":
+                            yield make_chunk(request_id, requested_model, f": `{repaired_args.get('path', '.')}`\n")
+                        elif func_name == "grep_search":
+                            yield make_chunk(request_id, requested_model, f": `{repaired_args.get('pattern', '')}`\n")
+                        else:
+                            yield make_chunk(request_id, requested_model, "\n")
+
+                        # Execute the tool
+                        result = await execute_tool(func_name, repaired_args)
+
+                        # Append teaching note if repaired
+                        if repair_notes:
+                            result += f"\n\n[SYSTEM REPAIR NOTE: The harness automatically fixed formatting issues: {', '.join(repair_notes)}. Please strictly follow the tool's JSON schema in subsequent calls without these wrapping/formatting errors.]"
+
+                        # Show truncated result to user
+                        preview = result[:500] + ("..." if len(result) > 500 else "")
+                        yield make_chunk(request_id, requested_model, f"```\n{preview}\n```\n")
+
+                        # Add tool result to message history
+                        final_messages.append({
+                            "role": "tool",
+                            "tool_call_id": tc["id"],
+                            "content": result,
+                        })
+
+                        await save_message(session_id, "tool", result, tool_call_id=tc["id"])
 
                 # Continue the agentic loop (model processes tool results)