Reviews

Kavi includes a built-in review system for examining agent-produced diffs. Reviews use structured dispositions rather than free-form comments, making it clear what action is needed for each piece of feedback.

Review Threads

A review thread is created against a diff produced by an agent task. Each thread contains:

• The diff being reviewed (file-level or hunk-level)
• One or more review comments with dispositions
• An assignee (agent or operator)
• Status tracking (open, resolved, auto-resolved)

Review threads are visible in the TUI and can be created by:

• The operator reviewing agent output
• An agent reviewing another agent's work (via peer messaging)
• Kavi automatically when potential issues are detected

Dispositions

Every review comment carries a structured disposition that indicates the nature of the feedback:

Approve

The change looks correct. No action needed. This is a positive signal that the reviewed code is acceptable.

Concern

Something may be wrong and warrants investigation. The assignee should evaluate the concern and either address it or explain why it's not an issue.

Question

A clarifying question about the change. The assignee should respond with an explanation.

Note

An informational comment that doesn't require action. Useful for documenting decisions or providing context for future reference.

Accepted Risk

The reviewer acknowledges a potential issue but accepts it as an intentional trade-off. This disposition documents known risks without blocking progress.

Won't Fix

The reviewer identifies an issue but determines it should not be addressed in the current scope. This is distinct from accepted risk -- it acknowledges the issue but defers it.

Assignees

Review comments can be assigned to:

• An agent (codex or claude) -- the agent receives the review feedback in its next prompt and is expected to address it
• The operator -- the feedback requires human judgment or action

When a review comment is assigned to an agent, Kavi includes the comment text, disposition, and relevant diff context in the agent's next prompt. This creates a feedback loop where agents can iterate on their work based on review input.

Follow-up Tasks

Review comments can generate follow-up tasks directly. When a concern or question results in actionable work, you can:

1. Select the review comment in the TUI
2. Create a follow-up task from it
3. The task is added to the current plan with a dependency on the reviewed task
4. The task includes the review context (comment, disposition, diff)

Follow-up tasks created from reviews are tagged with their origin, so you can trace from a review comment to the task that addressed it.

Auto-resolve

Review threads can auto-resolve when their parent task completes. This is useful for concerns that are addressed as part of normal task execution without explicit intervention.

Auto-resolve triggers when:

• The task associated with the review completes successfully
• The diff that prompted the review has been modified (the concern was addressed)
• The review disposition allows auto-resolution (concerns and questions, not accepted risks)

Auto-resolved threads are marked as such in the review history, distinguishing them from manually resolved threads.

Reply Threading

Review comments support reply threading, enabling back-and-forth discussion. Replies can come from:

• The operator responding to an agent's question
• An agent responding to a review concern
• Another agent providing additional context

Each reply is timestamped and attributed to its author, creating a conversation history within the review thread.

Review Workflow

A typical review workflow looks like:

1. Agent completes a task and produces a diff
2. Operator reviews the diff in the TUI
3. Operator adds review comments with appropriate dispositions
4. Comments assigned to agents are delivered in the next prompt
5. Agents address concerns and questions
6. Follow-up tasks are created for larger issues
7. Threads resolve (manually or automatically) as issues are addressed

Reviews are persisted in the session event log, providing a full audit trail of all feedback and responses.