Review

Generates a JSON review displayed in the TUI:

review "Context: Refactored authentication to use JWT tokens instead of sessions for better scalability"
review "Motivation: User requested real-time updates, so I added WebSocket support"
review "Background: Fixed the pagination bug where the last page wasn't loading correctly"

After the tool executes, simply acknowledge completion. Do not repeat the review content.

Example response:

Review complete. See the structured review above for details.

Text Mode

If the user prefers inline chat output instead of the TUI, use text mode:

review "Context: Major refactor to improve test coverage" --mode text

In text mode, the review is returned as markdown prose in the chat.

Parameters

• task (required): Context for the review. Provide background on why changes were made, design intent, or motivation. The narrator discovers what changed via git/jj. Do not list specific files or tell the narrator what to focus on unless the user explicitly requested a scoped review.
• mode (optional): Either "structured" (default, returns JSON for TUI) or "text" (returns markdown prose).

What to Pass in task

Good examples (context and motivation):

• "Context: Refactored auth to use JWT tokens for better scalability"
• "Motivation: User reported pagination bug on the last page"
• "Background: Added caching layer to improve performance"
• "Design intent: Simplified error handling to reduce boilerplate"

Bad examples (directives and file lists):

• "Review the authentication changes in src/auth/"
• "Check src/api/client.ts and src/types.ts"
• "Focus on the new API endpoints"

Exception: If the user explicitly asks for a scoped review ("review just the auth changes"), then it's appropriate to pass that scope in the task parameter.

The narrator agent discovers all changes via git/jj diff and decides what to include. Your job is to provide helpful context about why the changes were made, not what to look at.

Important Notes

• The narrator agent has its own context and uses read, grep, find, ls, and bash tools
• Reviews are generated from scratch each time - there is no persistent state
• The user sees the review immediately via TUI - never repeat the review contents back to them
• If the review fails, the error message will indicate what went wrong