A harness for long-horizon Claude Code work: goals decomposed into loop-runnable phases, hooks that catch the bad session, verified iterations, and decisions that stay yours.
Agent
Deep adversarial review of one task's diff against its brief/spec — the review half of subagent-driven development. Spawn per completed task (round 1) and per fix round. Verifies claims against artifacts, never trusts reports; read-only plus test execution.
Agent
Measures pattern-completeness across the whole codebase after pattern-introducing work — the structural countermeasure to "AI slop" (patterns applied to 30% of sites). Spawn before merging any branch that adds/changes a convention. Read-only; returns coverage counts per pattern.
Agent
Independent verification of a loop iteration's claimed progress. Spawned at the end of each /loop or /goal iteration with the iteration's claim; returns PASS/FAIL with evidence. Read-only plus test execution — never fixes anything itself.
Agent
Dedicated adjacent-risk sweep of a changed area — the structural substitute for "Fable-style noticing" on Opus/Sonnet. Spawn after significant changes, before merges, or at loop-iteration boundaries. Read-only; returns ranked load-bearing observations, not a lint report. Still valuable on Opus 5 as a dedicated…