MiniMax-AI / MiniMax-AI/minimax-code
Agent Plugin Skills cannot enforce parent tool restrictions during setup
Nobody has claimed this yet.
- Dominant language
- TypeScript
- Stars
- 1.3k
- Forks
- 141
- Avg merge
- 2h 45m
- Merged PRs (30d)
- 46
Description
Summary
Agent Plugin Skills currently cannot enforce a tool restriction on the parent MiniMax Code session while performing plugin-guided setup. This prevents certifying autonomous, native-only setup flows for security-sensitive plugins.
Environment
- MiniMax Code: 3.0.68.134
- Plugin type: Agent Plugin, imported from a public Git repository
- Setup initiated from a selected Plugin Skill in a disposable MiniMax profile
Reproduction
- Create or import an Agent Plugin with a Skill that guides a setup operation through MiniMax-native mechanisms (for example,
mavis agentprofile/configuration operations). - In the Skill prompt, explicitly provide all required input (including the temporary profile root) and explicitly prohibit Terminal, shell commands, HTTP requests, direct store edits, and direct agent workflow/system-prompt/project-file changes.
- Start the Skill from the MiniMax Code desktop UI.
- The Skill can invoke its MCP helper, but the parent session still has Terminal available and may invoke it (observed: a directory listing command) to inspect the plugin or profile before continuing.
Actual behavior
The parent agent's available tools are not constrained by the selected Skill's instructions. Prompt-level restrictions are advisory only: Terminal remains exposed and can be used during setup.
Custom Agent agent.md tool selectors can restrict child roles, but they do not restrict the parent session executing the Plugin Skill. Therefore they cannot provide an end-to-end guarantee for a native-only, shell-free setup flow.
Expected behavior
Please provide one of these host-enforced mechanisms:
- A declarative allowlist/denylist for tools available to the parent session while a Plugin Skill is running (for example, excluding Terminal and shell); or
- A non-prompt native setup API/surface for Agent Plugins to create/update supported agent configuration safely, without asking the parent model to perform file or shell operations.
The restriction must be enforced by the host, visible to the user, and testable by plugin authors.
Impact
Without a host-enforced boundary, a plugin cannot truthfully certify autonomous setup as native-only or shell-free. For production distribution, the only safe outcomes are a supervised manual setup contract or a fail-closed block. This affects plugins that need to install or configure multi-agent roles while preserving least privilege.
Could you confirm whether a parent-session tool policy already exists (and how Plugin Skills declare it), or whether this capability/API is planned?
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start by reproducing the setup flow from a selected Plugin Skill in the MiniMax Code desktop UI, observing the MCP helper and the parent session's Terminal access. Review how the Skill prompt and custom agent.md tool selectors are handled. Done means plugin authors have a host-enforced, user-visible, testable parent-session restriction or a native setup surface that prevents unauthorized shell or file operations.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- typescript
- Domain
- developer-experience, security, tooling
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Active
- Clarity
- Mostly clear
- Newbie friendliness
- 30/100