Context

I noticed something that seems inconsistent with how parallel tool calls should be handled. I'd appreciate clarification on whether this is intended behavior or a potential issue.

The Observation

In constructScratchPad, tool calls are grouped by comparing the Log field:

func (o *OpenAIFunctionsAgent) constructScratchPad(steps []schema.AgentStep) []llms.ChatMessage {
	// ...
	for i, step := range steps {
        if i == 0 || step.Action.Log != steps[i-1].Action.Log {
            // Start a new group
        }
    }
}

However, in ParseOutput, each tool call gets a unique Log that includes the tool name:

func (o *OpenAIFunctionsAgent) ParseOutput(contentResp *llms.ContentResponse) (
	[]schema.AgentAction, *schema.AgentFinish, error,
) {
    actions = append(actions, schema.AgentAction{
        Tool:      functionName,
        ToolInput: toolInput,
        Log:       fmt.Sprintf("Invoking: %s with %s %s", functionName, toolInputStr, contentMsg),
        ToolID:    toolCall.ID,
    })
}

The Issue: Even when multiple tool calls come from the same LLM response, each has a different Log (e.g., "Invoking: search..." vs "Invoking: calculator..."). This means step.Action.Log != steps[i-1].Action.Log is always true, causing each tool call to be treated as a separate group.

Example

When LLM returns parallel tool calls [search, calculator]:

Expected:

AI Message (ToolCalls: [search, calculator])
Tool Result (search result)
Tool Result (calculator result)

Actual:

AI Message (ToolCalls: [search, calculator])
Tool Result (search result)
AI Message (ToolCalls: [search, calculator]) ← Duplicate AI message
Tool Result (calculator result)