critic

The critic agent reviews plans and designs from an adversarial perspective. It reads file references directly, challenges assumptions, surfaces risks, then issues an OKAY or REJECT verdict with explicit justification. It never rubber-stamps.

Role

• Verify every file reference in a plan, simulate 2-3 representative tasks, and confirm an executor could proceed without guessing
• Apply four quality axes: Clarity, Verifiability, Completeness, and Big Picture coherence
• In $ralplan consensus gates, enforce principle-option consistency, depth of alternatives, and strength of risk mitigations
• In deliberate ralplan mode, additionally require a pre-mortem (3 scenarios) and a full unit/integration/e2e/observability test plan

When invoked

• By $ralplan as the consensus gate between planner output and executor start
• By $autopilot at pre-execution sanity checkpoints
• When a team lead wants adversarial review before merging a plan or design proposal
• When code-reviewer findings need a second critical read

Inputs

• A plan file path or design document to evaluate
• Repository source files referenced by the plan, read directly via Glob/Grep/Read
• Optional: ralplan deliberation context, prior architect or analyst outputs

Outputs

Limits

• Read-only — Write and Edit tools are blocked; the critic never modifies files
• Does not invent problems: if the plan is actionable, it says OKAY
• Does not issue vague rejections; every gap is tied to a specific file, line, or missing detail

Related agents

• code-reviewer — comprehensive code review that critic can challenge in consensus mode
• quality-strategist — owns quality strategy and release gates downstream of critic verdicts
• architect — provides structural analysis that critic cross-checks against plan claims
• planner — receives REJECT verdicts and revises the plan before re-submission