bug: injected field in completion message sent to generic openai api may cause pydantic validation errors #548
Description
Component
CLI
Summary
After updating from mistral-vibe 2.6.2 to 2.7.0. We get the following error on any request to our local llm api.
status: 400 Bad Request
reason: Bad Request
request_id: N/A
endpoint: http://mistral-devstral-2-25b.apps.localdomain.local/v1/chat/completions
mode: mistralai/Devstral-Small-2-24B-Instruct-2512
provider_message: 1 validation error for SystemMessage
injected
Extra inputs are not permitted [type="extra_forbidden, input_value=False, input_type=bool]
For further information visit htttps://errors.pydantic.dev/2.12/v/extra_forbidden
body_excerpt:{"error":{"message": 1 validation error for SystemMessage\ninjected\n Extra inputs are not
...
payload_summary: {"model":"mistrai/Devstral-Small-2-24B-Instruct-2512","message_count":2, "approx_chars"26943", "temperature":0.2, "has_tools":true, "tool_chıice:"auto"}
The problem appears to be from the extra addition of the injected field in the user/system messages sent to the completion api which seems to not to support it.
For now, I have managed to solve this by removing the injected field from vibe/core/types.py LLMMessage class or reverting to 2.6.2
This may be a problem on the api side, which forces a specific message model. Just wanted to report if someone else happens to have the same problem.
Reproduction steps
Run vibe to start and send any prompt to a model with:
api_style="openai"
backend="generic"
Versions / environment
mistral-vibe 2.7.0