Mistral Vibe: Custom Compaction Prompts, Token Budget Control, and Safer Programmatic Mode

Mistral Vibe

Mistral Vibe v2.11.1 introduced several improvements for developers who run the agent programmatically or operate it in long, context-heavy sessions. The most notable change is a security-minded default shift: programmatic mode (-p) no longer auto-approves tool calls by default, preventing unintended destructive actions in CI or scripted pipelines. Custom compaction prompts let teams define how context is summarized when the context window fills up, and a new --max-tokens flag enables hard token budget limits per session. Browser sign-in for Mistral providers is now enabled by default.

Sources & Mentions

2 external resources covering this update


Custom Compaction Prompts, Token Budget Control, and Safer Programmatic Mode

Mistral Vibe v2.11.1 brings several targeted improvements for teams running the agent programmatically or managing long, context-heavy sessions.

Programmatic Mode No Longer Auto-Approves Tool Calls (Breaking Change)

The most significant change in this release is a security default shift: programmatic mode (-p) no longer auto-approves tool calls by default.

Previously, running vibe -p in a CI pipeline or scripted environment would silently approve all tool calls β€” including destructive ones like file deletions or shell commands. This created a class of hard-to-debug production incidents where automated pipelines would approve actions that a human operator would have rejected.

With v2.11.1, programmatic mode requires explicit approval or a deliberate opt-in flag to restore the previous auto-approve behavior. Teams that rely on fully automated pipelines will need to update their invocations.

Custom Compaction Prompts

When the context window fills up, Vibe compacts the conversation history into a summary. Previously, the compaction format was fixed. Now, teams can define a compaction_prompt_id to control how this summarization is done.

Custom compaction prompts can be stored in two locations:

  • ~/.vibe/prompts/ β€” user-level prompts, available across all projects
  • .vibe/prompts/ β€” project-level prompts, checked into the repository

This is especially useful for codebases with domain-specific conventions: teams can instruct Vibe to preserve architectural decisions, key variable names, or ongoing task state across compaction boundaries.

--max-tokens Hard Token Budget Flag

A new --max-tokens flag lets operators set a hard ceiling on token consumption per session. Once the session reaches the limit, Vibe stops generating and reports the budget exhaustion β€” rather than silently continuing and incurring unexpected costs.

This is particularly useful in cost-sensitive CI environments or when running Vibe on metered API keys with strict spend limits.

Browser Sign-In for Mistral Providers Now GA

The enable_experimental_browser_sign_in flag has been retired. Browser-based authentication for Mistral providers is now enabled by default and exits experimental status. Teams that had explicitly set this flag can remove it from their configuration.


Mentioned onGitHubGitHub