Skip to content
MV

Mistral Vibe v3.1.4

Released · Mistral AI

Must-knowBreakingChanges background and headless runs, skills and plugins, permission rules, hooks, context and compaction settings

What the release notes say

Added

  • /mcp <source> has a search field that filters its tools by name and description
  • /thinking <level> sets the thinking level directly, without opening the picker
  • Each rail item opens a window in place of the input box: subagents, background tasks (s stops one), pull requests, the goal and its actions, and the todo list
  • collapse_turn_details = true collapses a settled turn's intermediate messages, reasoning and tool calls into a Worked for group above its final answer; a click or Ctrl+O expands it
  • keep_awake setting that keeps the machine from idle-sleeping while a turn runs or waits for your answer
  • Links in /release-notes open on click
  • Emacs Ctrl+P / Ctrl+N move to the previous / next line of the prompt (recalling history at its edges) and Ctrl+B / Ctrl+F move the caret by one character
  • Hovering a link to a local file previews its code in a popup
  • Clicking a zed:// editor link opens it in Zed
  • Copying a selection plays a short, subtle animation: a soft band sweeps across it
  • disable_selection_animation turns the copy animation off
  • /skills lists the skills that failed to load, with their file and the reason
  • /skills shows each skill's file, opens a local skill in your editor with e, and deletes it with x after a confirmation naming its folder
  • n in /skills starts a new skill with the skill creator
  • a in the /skills Shared tab adds the highlighted skill, and r removes a shared skill at once
  • --approval-policy and the default_approval_policy setting choose how tool calls are approved (ask, accept-edits, smart-approve, auto-approve), independently of the agent.
  • /agent lists your agents and switches to one, keeping the current approval policy.
  • Vibe is now also available as a desktop app: download it

Changed

  • The agent now explains what its questions are about, so they make sense without scrolling back
  • Local sessions now run on the Session API engine by default; set experimental_local_runtime_session_api = false (or VIBE_EXPERIMENTAL_LOCAL_RUNTIME_SESSION_API=false) to go back to the previous engine
  • /mcp keeps every server, connector and tool on one line: tool descriptions show dimmed after the name and end with an ellipsis, and narrow terminals drop columns before shortening names
  • Text typed while Vibe starts is now sent as soon as the session is ready when it ends with Enter, instead of waiting in the input
  • Vibe now trusts the operating system's certificates as well as its own by default, so it works behind corporate proxies without extra setup.
  • The agent asks about every decision it needs from you in one question dialog, no longer stopping at four
  • Subagents, background tasks, pull requests, the goal and the todo list now share one status rail of compact chips above the input box, replacing the goal and todo lines, the always-open subagent list and the footer's PR link
  • Waiting subagents stay in the subagent list instead of hiding behind a
  • The subagent list shows running subagents first and ready ones last
  • Press Ctrl+X then a letter, or click a chip, to open a rail item's window, even while another one is open
  • In the subagent and PR lists, Shift+Tab now moves to the previous rail window instead of cycling approval policies
  • The queue's Enter steer and Ctrl+C cancel hints moved from the loading line to the » Queued header
  • The tmux mouse off startup warning now hides with Ctrl+X then h, since Ctrl+X opens the status rail
  • The /branch and /fork panel now says its command opens the forked session
  • Dialogs, panels and onboarding screens now draw square border corners, matching the rest of the interface
  • Models that don't set temperature now use 1.0 instead of 0.2
  • Interrupting with Esc or Ctrl+C before Vibe answers takes the turn back and puts your prompt back in the input to edit and resend
  • Clarified task authorization and removed conflicting chat confirmation rules.
  • The scrollbar shows only its thumb, without a track.
  • /skills opens while the agent works; a change made during a turn applies when the turn ends
  • The /skills list shows the highlighted skill's description above its rows instead of its body, and the tabs show their counts and switch with Tab, Shift+Tab, ← and →
  • A skill's page in /skills opens on its description, then its status, source, scope and file, then its SKILL.md rendered as markdown; the arrows scroll it when there is no row to move to
  • My Skills in /skills is sorted by name, and the box keeps the height it opened with
  • A plugin's page in /plugins leads with its description, then its facts under dim labels, then its components by kind, and long paths wrap instead of being cut
  • Keyboard hints wrap onto more rows in a narrow terminal instead of being cut off, in every dialog, the loading line and the history search
  • Each session now runs an agent (default, lean, or a custom one) and an approval policy, changed independently: Shift+Tab cycles approval policies only, and the composer shows the agent next to the policy when it is not the default.
  • Old settings keep working: default_agent, --agent and saved sessions that name ask, accept-edits, smart-approve or auto-approve now set the approval policy.
  • An agent no longer decides what runs without asking: its safety, bypass_tool_permissions and permission = "always" settings are ignored, and a session runs your default_approval_policy whatever its agent. To keep a custom agent running in auto-approve, set default_approval_policy = "auto-approve" or pass --approval-policy auto-approve. An agent can still narrow a tool to ask or never. Other settings an agent may not set (such as a tool allowlist, sensitive_patterns, enabled_agents or enable_auto_approve) are ignored and listed as config issues.