Control the agent

Coding and planning

Coding mode exposes the normal enabled tool set under the active permission policy.

Planning mode changes instructions, tool availability, and enforcement. Research remains available; edit and write are constrained to Nerve’s plan directory; long-running task mutations and mutating Jira/Confluence operations are unavailable; shell commands pass a planning guard. A plan must be presented for review before implementation.

Planning restrictions reduce accidental mutation but are not an operating-system sandbox.

Permission levels

• Read-only: permits local read tools. It blocks edit/write, Bash, Python, ordinary network calls, Jira/Confluence network reads, and Explore child spawning.
• Supervised: asks for approval on non-read risks. You can choose whether selected local read tools and task status/log inspection autoapprove.
• Autonomous: executes allowed risks without normal approval pauses.

Policy is evaluated for every dispatch and recorded with the tool lifecycle.

Model and thinking

Select an authenticated model and one of its declared thinking levels. Nerve knows seven possible levels from off through max, but the picker only presents supported choices and runtime resolution can clamp a stale request.

Changing model, thinking, mode, or permission on an active conversation does not recreate it. Changes made during a run apply to subsequent work—especially the next provider request—not the request already streaming.

Choosing a baseline

Use read-only for investigation where no network or child agents are needed. Use supervised for unfamiliar work and most first runs. Reserve autonomous for bounded tasks in a trusted, recoverable workspace.

Next steps

• Approvals, questions, and plans
• Security model
• Tool reference