Fix tool calls not being persisted when tools error

[a-rebets]

Problem:

When a tool execution throws an error, the assistant message containing the tool-call is not saved to the database. Only the tool-result message with the error is persisted, breaking the message chain.

Root Cause:

In src/mapping.ts, the serializeNewMessagesInStep function uses step.toolResults.length to determine which messages to save:

const messages: MessageWithMetadata[] = await Promise.all(
  (step.toolResults.length > 0
    ? step.response.messages.slice(-2)  // assistant + tool messages
    : step.content.length
      ? step.response.messages.slice(-1)  // Only last message
      : [{ role: "assistant" as const, content: [] }]
  ).map(...)
);

When a tool fails:

• step.toolResults is empty (no successful execution result)
• step.response.messages contains both the assistant message with tool-call AND the tool message with error
• The code incorrectly takes only the last message, missing the assistant message

What gets saved to the database (example):

[
  {
    content: "list the datasets",
    role: "user",
  },
  // MISSING: assistant message with tool-call
  {
    content: [
      {
        output: {
          type: "error-text",
          value: "Your request couldn't be completed. Try again later.",
        },
        toolCallId: "call_somehash",
        toolName: "datasets_listDatasetsTool",
        type: "tool-result",
      },
    ],
    role: "tool",
  },
]

After all of this we see warnings like:

[WARN] 'Tool result without preceding tool call.. adding anyways'

Proposed Fix:

Check if there are tool messages in step.response.messages instead of relying solely on the length

---

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[a-rebets]

Could be related to #162

[a-rebets]

Adjusted the fix, can confirm in my DB the history is now saved properly:

[
	{
	  content: "so what datasets do I have?",
	  role: "user",
	},
	{
	  content: [
	    {
	      text: `**Executing dataset tool**
	
	I need to call the datasets_listDatasetsTool since this request is actionable. I’ll make sure to ...`,
	      type: "reasoning",
	    },
	    {
	      args: {},
	      toolCallId: "call_3GrNAa0RsNexZ7F3D2frl1kN",
	      toolName: "datasets_listDatasetsTool",
	      type: "tool-call",
	    },
	  ],
	  role: "assistant",
	},
	{
	  content: [
	    {
	      output: {
	        type: "error-text",
	        value: "This is a test!!!",
	      },
	      toolCallId: "call_3GrNAa0RsNexZ7F3D2frl1kN",
	      toolName: "datasets_listDatasetsTool",
	      type: "tool-result",
	    },
	  ],
	  role: "tool",
	},
	{
	  content: [
	    {
	      text: `**Clarifying tool response**
	
	The tool returned an unusual message: "This is a test!!!"...`,
	      type: "reasoning",
	    },
	    {
	      text: `I checked the saved datasets — there are no usable datasets found.`,
	      type: "text",
	    },
	  ],
	  role: "assistant",
	}
]

[ianmacartney]

in the case of error, is the finishReason tool-calls?
My understanding of tool-calls finishReason was that it was the signal that there were tool calls that need response, e.g. when you have a tool with no execute function, it will finish with tool-calls but there will only be one response message - the assistant response.