Native Mistral API fails with `cache_breakpoint` validation error during Agent execution

Description

When using the native Mistral API (api.mistral.ai) with CrewAI, standalone LLM calls work correctly, but Agent execution fails.

The same LLM configuration successfully returns a response using:

llm.call("Say hello")

However, when the LLM is attached to a CrewAI Agent and executed through crew.kickoff(), the request fails with:

litellm.BadRequestError: MistralException -

{
  "detail": [
    {
      "type": "extra_forbidden",
      "loc": ["body","messages",0,"system","cache_breakpoint"],
      "msg": "Extra inputs are not permitted"
    },
    {
      "type": "extra_forbidden",
      "loc": ["body","messages",1,"user","cache_breakpoint"],
      "msg": "Extra inputs are not permitted"
    }
  ]
}

This appears to occur only in the Agent execution path (call_llm_native_tools), not during direct LLM usage.

Steps to Reproduce

1. Create a CrewAI LLM using:

llm = LLM(
model="mistral/mistral-large-latest",
api_key=os.getenv("MISTRAL_API_KEY")
)

2. Verify this works:

print(llm.call("Say hello"))

3. Create an Agent using the same LLM.

4. Execute a simple Crew with crew.kickoff().

5. Observe that Agent execution fails with the cache_breakpoint validation error from the Mistral API.

Expected behavior

Agent execution should work the same way as direct LLM.call() and should not send unsupported fields to the native Mistral API.

Screenshots/Code snippets

Include both the minimal reproduction code and the relevant error.

from crewai import LLM, Agent, Task, Crew, Process
from dotenv import load_dotenv
import os

load_dotenv()

llm = LLM(
model="mistral/mistral-large-latest",
api_key=os.getenv("MISTRAL_API_KEY"),
)

print(llm.call("Say hello"))

agent = Agent(
role="Researcher",
goal="Research a topic",
backstory="Research assistant",
llm=llm,
)

task = Task(
description="Say hello",
expected_output="Greeting",
agent=agent,
)

crew = Crew(
agents=[agent],
tasks=[task],
process=Process.sequential,
)

crew.kickoff()

Error

litellm.BadRequestError: MistralException -

{
  "detail": [
    {
      "type": "extra_forbidden",
      "loc": ["body","messages",0,"system","cache_breakpoint"],
      "msg": "Extra inputs are not permitted"
    },
    {
      "type": "extra_forbidden",
      "loc": ["body","messages",1,"user","cache_breakpoint"],
      "msg": "Extra inputs are not permitted"
    }
  ]
}

Operating System

Windows 11

Python Version

3.12

crewAI Version

1.15.10

crewAI Tools Version

1.15.10

Virtual Environment

Venv

Evidence

Possible Solution

It appears that CrewAI's Agent execution path (call_llm_native_tools) injects a cache_breakpoint field into the messages payload before forwarding the request to LiteLLM.

The native Mistral API strictly validates the request schema and rejects unknown fields with an extra_forbidden error.

A possible fix would be to:

• Avoid sending cache_breakpoint to providers that do not support it (such as the native Mistral API), or
• Strip provider-specific metadata before making the API request, or
• Add a provider compatibility check so unsupported fields are not included in requests to Mistral.

Additional context

I verified that this is not a Mistral API key or model configuration issue.

The following direct call succeeds:

llm = LLM(
    model="mistral/mistral-large-latest",
    api_key=os.getenv("MISTRAL_API_KEY")
)

print(llm.call("Say hello"))

Start at the Agent execution path named call_llm_native_tools and compare its message payload with the direct LLM.call() path. Reproduce the issue with the supplied Mistral configuration and verify that crew.kickoff() no longer sends cache_breakpoint fields to the native Mistral API while direct calls continue to work.