Separate the layers before configuration
Installing another extension is not a substitute for understanding which layer is missing.
- Agent provider
- The task-running agent selected in the IDE.
- Model
- The model used for inference, which may be a separate choice.
- MCP server
- An external tool and data endpoint that requires its own permissions.
- Skill or repository instruction
- Procedural guidance the agent may read and apply.
- Plugin or hook
- A package or execution extension; inspect its code and access independently.
A low-risk first session
- Check the current GitHub Copilot and JetBrains documentation and install the documented Codex CLI prerequisite if required.
- Select the provider and verify the exact local project, branch and files the IDE exposes.
- Start with a read-only question about a known file; compare the answer with the file.
- Check permission prompts and logs for the tools actually called.
- Try a small reversible edit on an isolated branch or worktree.
- Review the diff and run the repository checks before accepting it.
Trace authority to the right component
A provider may be available while a particular MCP server is not. A tool may be connected but lack the required scope. Keep credentials and broad filesystem access out of the trial, and do not infer that a skill grants a tool capability.
If the provider, model or tool source is unclear, stop at read-only analysis and inspect the client configuration.