Враждебный аудит кодовой базы параллельными агентами

testing

audit-adversarial is a Claude Code skill that performs a hostile full-codebase review of Caffeine by spawning parallel subagents that receive no design context whatsoever — no design-decisions.md, no synchronization.md, no .claude/docs. Each of the 4–6 reviewers covers a distinct subsystem (core, guava, jcache, simulator, examples) and probes eleven attack surfaces: memory model violations, state-corruption interleavings, resource leaks under OOME/SOE, silent data loss, ConcurrentMap contract breaches, O(n) operations on O(1) paths, sentinel collisions, and more. A second layer of evaluator subagents challenges every finding or zero-finding claim, forcing reviewers to defend conclusions with concrete interleavings or drop them. Results are consolidated, deduplicated, escalated to dynamic tools like JCStress when static analysis is insufficient, and finally adjudicated against design docs to distinguish confirmed bugs from intentional trade-offs. The skill is built for teams who need an adversarial perspective unfiltered by maintainer familiarity.