Mastra•10mo ago

Unsupported role: tool when using memory

When calling tools on the client and using

useChat

useChat

useChat

useChat

onToolCall()

onToolCall()

onToolCall()

onToolCall() callback, i return the result, however on my mastra server i get a:
MessageConversionError [AI_MessageConversionError]: Unsupported role: tool

this happens on

 "@mastra/core": "0.6.1-alpha.0"

"@mastra/core": "0.6.1-alpha.0"

 "@mastra/core": "0.6.1-alpha.0"

"@mastra/core": "0.6.1-alpha.0" and

 "@mastra/core": "0.6.1-alpha.1"

"@mastra/core": "0.6.1-alpha.1"

 "@mastra/core": "0.6.1-alpha.1"

"@mastra/core": "0.6.1-alpha.1".

and it only happens while using memory, here is my exact implementation of the memory:

 memory: new Memory({
    storage: new LibSQLStore({
      config: {
        url: process.env.DATABASE_URL ?? 'file:memory.db',
      },
    }),
    options: {
      workingMemory: {
        enabled: true,
        use: "tool-call", // Required for toDataStream() compatibility
      },
    },
  }),

 memory: new Memory({
    storage: new LibSQLStore({
      config: {
        url: process.env.DATABASE_URL ?? 'file:memory.db',
      },
    }),
    options: {
      workingMemory: {
        enabled: true,
        use: "tool-call", // Required for toDataStream() compatibility
      },
    },
  }),

 memory: new Memory({
    storage: new LibSQLStore({
      config: {
        url: process.env.DATABASE_URL ?? 'file:memory.db',
      },
    }),
    options: {
      workingMemory: {
        enabled: true,
        use: "tool-call", // Required for toDataStream() compatibility
      },
    },
  }),

 memory: new Memory({
    storage: new LibSQLStore({
      config: {
        url: process.env.DATABASE_URL ?? 'file:memory.db',
      },
    }),
    options: {
      workingMemory: {
        enabled: true,
        use: "tool-call", // Required for toDataStream() compatibility
      },
    },
  }),

the memory works though

OmicrxnOP•3/16/25, 5:28 PM

i have to say that if i remove memory the tool calling works, this only happens if i use memory

rase-•3/17/25, 5:50 AM

@Tyler would probably be able to help with this !

Ward•3/17/25, 8:49 AM

I'll have a quick look as well

OmicrxnOP•3/17/25, 1:06 PM

this was fixed a couple of days ago, which is where we found the [messages.at(-1)] thing but seems to be broken again, also sometimes it calls the tool a lot of times

Tyler•3/17/25, 4:17 PM

@Omicrxn when it calls the tool multiple times that's from before the bug fix so I wonder if there was a bad publish or it's pulling in an old version somehow

Tyler•3/17/25, 4:18 PM

Even with the fix the multiple tool calling thing can happen if more than one message is being sent in - though from what you said that's not happening

Tyler•3/17/25, 4:19 PM

Unsupported role: toolUnsupported role: tool is odd

OmicrxnOP•3/17/25, 4:29 PM

Yeah i have all the measures you told me from the last bug fixing applied, this started happening when updating to the newest 0.6.1-alpha.0 and .1

OmicrxnOP•3/17/25, 4:30 PM

The unsupported role is new though

Tyler•3/17/25, 5:08 PM

Thanks, I'll take a look - really strange since I don't think anything changed since then

OmicrxnOP•3/17/25, 6:18 PM

I'll try deleting node modules but I think i already tried. It's weird.

Tyler•3/17/25, 6:30 PM

Yeah I'm sure it's not on your side - we're currently planning out OSS stability+tests improvements but it'll be a week or two before we see those improvements. There's a good chance this is a regression or new bug

OmicrxnOP•3/17/25, 7:01 PM

No problem, I totally get the stage you are at, you have to iterate quick which you are doing an amazing job at. And writing tests would be beneficial but not the topmost priority it's okay

Tyler•3/17/25, 8:00 PM

being able to ship fast and not break anything is one of our main focuses right now so we'll be putting a huge focus on it this week and the next few

OmicrxnOP•3/17/25, 9:14 PM

hey @Tyler how can i get the [messages.at(-1)] on the stream on a mastra server {MASTRA_BASE_URL}/agents/:agentId/stream

OmicrxnOP•3/17/25, 9:15 PM

i'm just testing some things using a newly created mastra server just in case my implementation is at fault but I don't really know how to do that

OmicrxnOP•3/17/25, 9:15 PM

i've tried using the mastra client-js on my svelte project on a /api/chat route so that useChat goes there but instead of vercel it finds the mastra agent, but the client-js stream does not have a toDataStreamResponse()

Tyler•3/17/25, 9:16 PM

Good question! I'm not 100% sure off the top of my head. What does your client code look like?

OmicrxnOP•3/17/25, 9:23 PM

export const POST: RequestHandler = async ({ request }) => {
  const { messages, threadId, resourceId } = await request.json();

  const agent = mastraClient.getAgent(AGENT_NAME);
  const result = await agent.stream({
    messages: [messages.at(-1)],
    threadId,
    resourceId,
  });

  return result;
};

export const POST: RequestHandler = async ({ request }) => {
  const { messages, threadId, resourceId } = await request.json();

  const agent = mastraClient.getAgent(AGENT_NAME);
  const result = await agent.stream({
    messages: [messages.at(-1)],
    threadId,
    resourceId,
  });

  return result;
};

OmicrxnOP•3/17/25, 9:23 PM

as a suggestion maybe the client-js API could be similar to the mastra/core one, like stream(messages,{params}) instead of stream({params})

OmicrxnOP•3/17/25, 9:24 PM

and the same with the response not having toDataStreamResponse etc but this i'm not 100% sure if maybe it's okay because it has a different purpose, just random thoughts

Tyler•3/17/25, 9:30 PM

ah so the client-js is meant to be called client side. If you're in a server endpoint like that you should be able to import your agent and use it directly

OmicrxnOP•3/17/25, 9:31 PM

the thing is that my agent lives on a separate mastra server made with mastra create

OmicrxnOP•3/17/25, 9:32 PM

but since i don't have access to the stream endpoint and i can't use the experimental_previewRequestBody in svelte i don't know how to pass only the last message

OmicrxnOP•3/17/25, 9:34 PM

or maybe the created mastra streaming endpoint already does the [messages.at(-1)] thing?

Tyler•3/17/25, 10:48 PM

I don't think it does since passing just the last message is needed because of how useChat sends all messages each time. I haven't tried using it in your exact setup though so I'll give it a shot and see what's up

OmicrxnOP•3/17/25, 11:09 PM

perfect the setup is a mastra server mastra createmastra create and mastra devmastra dev and a useChat instance in svelte with api:'localhost:4111/api/agent/:agentID/streamapi:'localhost:4111/api/agent/:agentID/stream

OmicrxnOP•3/17/25, 11:11 PM

then the memory using working memory:

new Memory({
      options: {
        workingMemory: {
          enabled: true,
          use: "tool-call", // Required for toDataStream() compatibility
        },
      },
  })

new Memory({
      options: {
        workingMemory: {
          enabled: true,
          use: "tool-call", // Required for toDataStream() compatibility
        },
      },
  })

and the issue is that although it recalls correctly the memory, when calling tools, they enter in a loop, and also the error at the beginning of the thread.

OmicrxnOP•3/17/25, 11:11 PM

I also tried witha custom endpoint where i could do the [messages.at(-1)][messages.at(-1)] and the same happens (which is when i opened this thread)

OmicrxnOP•3/17/25, 11:12 PM

I can work without memory for now but I plan on launching soon and I think memory would be really great on launch so let me know how it goes.

OmicrxnOP•3/17/25, 11:13 PM

this is an example of it calling multiple tools

OmicrxnOP•3/17/25, 11:14 PM

hope it helps!

Tyler•3/17/25, 11:23 PM

Thanks, that's helpful context - I'll keep you updated

Original message was deleted

OmicrxnOP•3/18/25, 10:23 PM

hey @Tyler is this fixing the issue on this thread?

Tyler•3/18/25, 10:23 PM

@Omicrxn unfortunately no, it's a perf improvement @rase- made

OmicrxnOP•3/21/25, 12:41 PM

hey @Tyler how are we on this?

Tyler•3/21/25, 6:50 PM

Hey @Omicrxn , lots going on this week so havn't had time to look into it! Have it on my list to check out today though

Tyler•3/21/25, 9:15 PM

@Omicrxn I took a look but it's working for me

wish I could reproduce it! Just pushed up another commit to my svelte repo, lmk if I'm missing anything that you have in your setup https://github.com/TylerBarnes/svelte-usechat/commit/83fd957f2060c48f8d7d727630b1e2c1ca6b4afc

GitHub

latest repro · TylerBarnes/svelte-usechat@83fd957

Tyler•3/21/25, 9:15 PM

OmicrxnOP•3/22/25, 1:14 PM

Hey @Tyler I have copied exactly your code into mine, but i'm still getting the:
Error streaming from agent MessageConversionError [AI_MessageConversionError]: Unsupported role: tool

OmicrxnOP•3/22/25, 1:15 PM

i have the same @mastra/* versions, the same setup, same usechat and client and server

OmicrxnOP•3/22/25, 1:16 PM

and in this example i just said my name which successfully called the updateMemory tool , then i closed the server and the app, init again and after that it gives this error

OmicrxnOP•3/22/25, 1:16 PM

just by saying hello

OmicrxnOP•3/22/25, 1:16 PM

didn't even called a tool yet

OmicrxnOP•3/22/25, 1:18 PM

I even tried using your repo but with my server and it gives the same error of unsupported tool so it is 100% on my mastra server

OmicrxnOP•3/22/25, 1:20 PM

and in the server the only different thing between yours and mine is that i don't use the embedder from openai:
my agent:

import { openai } from '@ai-sdk/openai';
import { Agent } from '@mastra/core/agent';
import { clientTools } from '../../tools';
import { Memory } from '@mastra/memory';

const instructions = `... too long for discord`;

const memory = new Memory({
  options: {
    workingMemory: {
      enabled: true,
      use: "tool-call", // Required for toDataStream() compatibility
    },
  },
});

export const uAgent = new Agent({
  name: "U",
  instructions: instructions,
  model: openai("gpt-4o-mini"),
  tools: clientTools,
  memory
});

import { openai } from '@ai-sdk/openai';
import { Agent } from '@mastra/core/agent';
import { clientTools } from '../../tools';
import { Memory } from '@mastra/memory';

const instructions = `... too long for discord`;

const memory = new Memory({
  options: {
    workingMemory: {
      enabled: true,
      use: "tool-call", // Required for toDataStream() compatibility
    },
  },
});

export const uAgent = new Agent({
  name: "U",
  instructions: instructions,
  model: openai("gpt-4o-mini"),
  tools: clientTools,
  memory
});

OmicrxnOP•3/22/25, 1:35 PM

@Tyler Here is a video of me using your repo, but with my mastra server

OmicrxnOP•3/22/25, 2:23 PM

Tried with the openai embedder but same result

OmicrxnOP•3/22/25, 4:55 PM

removing the workingMemory, seems to work fine, any insights on why? @Tyler

const memory = new Memory({
  options: {
    // workingMemory: {
    //   enabled: true,
    //   use: "tool-call", // Required for toDataStream() compatibility
    // },
  },
});

const memory = new Memory({
  options: {
    // workingMemory: {
    //   enabled: true,
    //   use: "tool-call", // Required for toDataStream() compatibility
    // },
  },
});

Unsupported role: tool when using memory

Similar Threads

Similar Threads

Similar Threads