Back to rankings

Mistral Vibe: Custom Subagent Permissions, Tool Rules and Skill Limits Now Enforced

Mistral VibeView original changelog

Mistral Vibe 3.0.2 fixed a gap where custom subagents ignored the restrictions their owners set. A subagent's [tools.*] permission, allowlist and denylist now apply to its tool calls, and its enabled_skills and disabled_skills now limit which skills it can load. Hooks from hooks.toml, including post_agent hooks, now also run for tool calls made inside subagents.

Key Takeaways

  • Custom subagent tool rules are now enforced, so permission, allowlist and denylist entries in [tools.*] actually limit what a subagent can do.
  • Subagent rules apply on top of the main agent's rules, so a subagent cannot be more permissive than the agent that spawned it.
  • Skill scoping works: enabled_skills and disabled_skills now restrict which skills a subagent may load.
  • Hooks now fire inside subagents, including post_agent hooks, so audit and policy hooks are no longer bypassed by delegation.
  • Before 3.0.2 these settings were silently ignored, so existing subagents may now behave more restrictively.
  • The change matters most for teams that rely on subagents for safe delegation, where an unenforced restriction was a real isolation gap.

Subagent restrictions now actually apply

Mistral Vibe 3.0.2 closes a trust-boundary gap in custom subagents. Before this release, several controls that a user or team wrote into a custom subagent's configuration were silently ignored. A subagent could therefore act more broadly than its configuration said it should. Version 3.0.2 makes those settings take effect.

Tool permissions, allowlists and denylists

A custom subagent's [tools.*] settings for permission, allowlist and denylist now apply to its tool calls. They apply on top of the main agent's own rules instead of being skipped. A subagent restricted to read-only tools, or blocked from certain commands, is now held to that restriction. The main agent's rules still apply as well.

Skill limits

A custom subagent's enabled_skills and disabled_skills now limit which skills the subagent can load. Previously the lists were ignored. Teams that give a narrow subagent only the skills it needs can now rely on that scoping.

Hooks run inside subagents

Hooks defined in hooks.toml now run for tool calls made inside subagents, including post_agent hooks, just as they do for the main agent. Hooks are commonly used for auditing, policy checks and formatting. Before this fix, delegating work to a subagent could skip those checks.

Why it matters

Delegation to subagents is a core way to use Mistral Vibe on larger tasks. Anyone who configured a subagent with tighter rules than the main agent should know these settings were previously not enforced, and that they are enforced from 3.0.2 on. Behavior that used to slip through may now be blocked or prompt for approval.