Mistral Vibe v2.15.0 Changelog

This release is mostly about making the agent feel smoother and less interruptive while you work. If you use Mistral Vibe day-to-day, you'll notice fewer annoying pauses, a cleaner interface, and more control over what happens behind the scenes. Power users get an early look at hooks — custom scripts that can intercept and shape what the agent does on every tool call.

What's new

• Message queue: If you type while the AI or a shell command is still running, your messages line up above the input box instead of vanishing. Press Esc to pause the queue, Ctrl+C to drop the most recent queued message, and Enter to send everything when paused.
• Experimental hooks: You can now write shell scripts in a hooks.toml file that run before and after every tool call. A hook can block a call, change what gets sent to the tool, or add extra context to the result. Turn this on with enable_experimental_hooks = true.
• Collapsed tool results: Tool output is folded up by default so you can scan a conversation without walls of text getting in the way. For web tools, the collapsed view still shows the URL and search query at a glance.
• No approval needed for safe commands: Common read-only shell commands like ls, cat, and pwd are now allowed automatically, so you're not asked to approve every harmless lookup.
• Delete sessions from the resume picker: You can remove an old session right from the list where you choose which one to reopen.
• Per-server MCP authentication: You can now set up login details for each MCP server individually using an [mcp_servers.auth] block in your config. MCP (Model Context Protocol) is a way for the agent to talk to external tools and data sources.
• max_turns over ACP: You can now set a limit on how many back-and-forth turns the agent takes, communicated through ACP (Agent Communication Protocol) via set_config_option.

Improvements

• Compaction keeps your goals intact: When the conversation gets too long and the system compresses older context, it now re-injects your earlier messages so the agent doesn't forget what you originally asked for.
• Model refusals are visible: If the AI model refuses to answer something, you now see that reason instead of the session just going silent with no explanation.

Fixes

• Blank screen on app switch: The CLI interface no longer wipes itself blank when you switch away and back.
• Messages lost during compaction: User messages are now correctly preserved even when context compression happens multiple times.
• Weird text input: Non-UTF-8 characters (text that doesn't follow the standard encoding) no longer break things across the CLI.
• Resume from risky folders: --resume now shows your history properly even when launched from a directory flagged as dangerous.
• Startup crashes: The app no longer crashes when starting in folders you don't own or when a parent git folder can't be read.
• Web search crash: Plain-text answers from web_search are handled without crashing.
• Mouse code leaks: Garbled mouse tracking data (NaN SGR reports) no longer spills into your chat input.

Breaking changes

• Experimental hooks retry behavior changed: If you were using post_agent_turn hooks, they no longer trigger a retry by exiting with code 2. Instead, a hook must exit 0 and print a JSON line like {"decision": "deny", "reason": "..."} to stdout. Exiting with code 2 is now treated as a plain failure. This only affects the experimental hooks feature.