Federico Del Gaudio
Norn docs AI Assist Contents
Back to project
•
3 min read

AI Assist

Table of Contents

Norn integrates Claude and Codex into the review workflow without turning them into automatic publishers. The assistant proposes; the reviewer decides what to keep and what to publish.

How it works

  • A run starts with the context of the current repository, pull request, and diff.
  • Norn stores review results, findings, and evidence locally.
  • Proposed comments remain drafts until you explicitly choose to publish them.
  • You can configure the provider and model options for the review workflow.

Norn launches the provider’s locally installed CLI. Authenticate claude or codex before starting a review, and use norn doctor --machine-only to check that the selected executable is available.

Reviews from the terminal

Select Codex explicitly:

norn review --repo-path . --scope working-tree \
  --ai-provider codex --allow-provider-diff

Or use Claude:

norn review --repo-path . --scope working-tree \
  --ai-provider claude --allow-provider-diff

You can request structured output for automation and integrations:

norn review --repo-path . --scope branch --format json \
  --fail-on-findings --allow-provider-diff

Use --model and --effort for one-run overrides, or configure defaults in the desktop app and TUI settings.

Privacy and control

For a non-empty headless review, --allow-provider-diff grants one-run consent to send only the selected diff and review instructions to the chosen provider. You can persist or revoke that local choice with:

norn setup --allow-provider-diff --yes
norn setup --deny-provider-diff --yes

Diff-sharing consent is separate from permission to run outside a coding agent’s sandbox. Codex and Claude Code must grant the exact norn review process enough host access to reach the provider CLI, its configuration, the OS credential store, and the network. Norn never edits or broadens those agent permissions.

Norn keeps credentials outside repository config and never publishes model output automatically. Review every finding and draft before publication.

Headless mode returns findings only. Desktop and TUI workflows let you turn accepted findings into local drafts and explicitly publish them.

When to use it

AI Assist is useful for navigating large pull requests, identifying risks, and getting an initial reading of unfamiliar code. It does not replace reviewer judgment: always verify every finding against the diff and the system’s actual behavior.