#90

Add Mistral Vibe as a supported agent harness

Openbenjamincanacopened 8mo ago

Mistral recently released Vibe, their open-source CLI coding agent (similar to Claude Code, Codex, and Gemini CLI). It supports a programmatic mode out of the box:

vibe --prompt "task description" --auto-approve --max-turns 50 --output json

It would be great to have it as a first-party agent harness in agent-eval, so Devstral (and future Mistral models) can be tested with their native agent rather than relying on OpenCode as a fallback.

Key details:

  • Repo: https://github.com/mistralai/mistral-vibe
  • Install: uv tool install mistral-vibe or pip install mistral-vibe
  • Auth: MISTRAL_API_KEY env var
  • Programmatic mode: vibe --prompt <task> --auto-approve with --max-turns, --max-price, and --output json options
  • Transcript support: Supports --output streaming for newline-delimited JSON per message, which could be used for transcript parsing

Activity

1
gaojude7mo ago

Feel free to PR this!