reviewer_mode
Scope: domain • Default: llm_with_deterministic_fallback
How review is performed for tasks in the review status.
Positions
| Position | Behavior |
|---|---|
llm_only | A reviewer LLM (or fan-out of persona reviewers) produces the verdict. Fails hard if the LLM is unavailable. |
deterministic_only | Only rule-based checks (lint, typecheck, rubric heuristics) run. No LLM cost. |
llm_with_deterministic_fallback | LLM reviewers first; if they timeout or error, deterministic checks provide the verdict. |
Cost vs confidence
- LLM-only: highest-quality verdicts; highest cost; single point of failure if the provider is down.
- Deterministic-only: cheap; fast; blind to anything a rubric can't encode (tone, clarity, architecture).
- Fallback: best of both.
Related
reviewer_fanout_policy— how multiple LLM reviewers are aggregated whenreviewer_modeis LLM-based.