Reviewer Agent

The Reviewer agent role, multi-criteria scoring, and revision control.

The Reviewer Agent (◆)

The Reviewer acts as the senior software engineering lead on the pipeline. It evaluates the combined output of the Planner, Coder, and Tester before any code is committed.

Evaluation Criteria & Scoring

The Reviewer scores the change across 4 core dimensions on a scale from 1 to 10:

1. Correctness (1–10): Does the code accurately satisfy all requirements without introducing regressions?
2. Code Quality (1–10): Is the code idiomatic, clean, maintainable, and conformant with the project’s conventions?
3. Test Coverage (1–10): Are edge cases, error conditions, and happy paths adequately verified?
4. Spec Adherence (1–10): Did the Coder follow the constraints defined in the TaskSpec?

Verdicts & Revision Loops

The Reviewer issues one of three verdicts in its ReviewVerdict artifact:

• Approved: The Reviewer accepts the change. NIKI ends the revision loop; a branch is still created only for a non-empty diff with no blocked test failure.
• RevisionNeeded: Issues were discovered. The Reviewer provides actionable feedback and starts another Coder round, bounded by general.max_revision_rounds.
• Rejected: The Reviewer ends the loop with the rejection recorded in report.md. Rejection alone does not create or block a branch; branch eligibility still depends on diff and verification gates.

[agents.reviewer]
provider    = "anthropic"
model       = "claude-sonnet-4-20250514"
max_tokens  = 8192
temperature = 0.1