Skip to content

Review approvals, questions, and plans

Nerve places review gates beside the work that caused them. While one is pending, the normal composer is disabled so a new prompt cannot bypass the decision.

In supervised mode, non-read actions pause before execution. Review the tool name, arguments, risk, workspace target, and any preview. Approve only the request you understand. A denied or failed action remains visible in the transcript.

Read-only tools can also require approval when read autoapproval is disabled. Autonomous permission skips normal approval for allowed risks; read-only policy denies disallowed tools rather than offering an approval override.

The question card accepts a free-text reply and provides common quick replies. It can also use the shared voice-input session. Dismiss when the question is no longer relevant.

A planning run writes a reviewed plan in Nerve plan storage and presents it in the transcript. Acceptance options can:

  • implement in the same conversation;
  • compact context, then implement;
  • optionally open a new implementation conversation and choose its model.

Review file scope, validation, migration impact, and security assumptions before accepting. A plan is a proposal, not evidence that implementation is already complete.

Tool lifecycle states—requested, started, output, completed, or failed—remain in history. Replayed events are deduplicated after reconnect, and long outputs are bounded or stored as artifacts/transcripts rather than silently injected in full.