Mighty Reviewer

Triggers after a coding turn finishes: it checks that an editing tool actually ran and that the git diff changed since the start of the turn, then spawns a background child session to review it so the review never clutters the conversation. Mechanical gates run first, the project's own typecheck, lint, and a scan for leftover debug output, with any failure counted as an automatic P1 finding. Four critic agents then run in parallel: adversarial-risk-critic, design-principles-critic, security-checklist-critic, and ai-slop-critic, each scoped to a different class of problem.

Every P0/P1 finding is adversarially verified before it is accepted: the review session reads the full file, checks imports and callers, and drops anything it cannot confirm. Critics carry their own false-positive controls, a confidence threshold, a rule against flagging failures with no concrete scenario, and a cap on how aggressively the AI-slop critic alone can block a ship, and they run read-only with edit, write, patch, and task disabled by mechanism. On NO-SHIP, the findings report is injected back into the coding session as a new prompt, triggering a fresh review, capped at two re-review cycles and ten injections per session so the loop cannot run forever.

On OpenCode 2, a TUI companion module adds live toasts on each review phase, a footer line showing progress under the prompt box, and a collapsible sidebar panel with a row per critic (severity counts, token totals, and a verified-finding total once the review finishes). Published on npm, it supports OpenCode 1.18.29+ and OpenCode 2.x from the same package, with all four critic agents registered by the plugin itself.