dsh-review

The biggest problem with AI code review is not missing bugs but reporting too many that are wrong. Run a crowd of parallel reviewers over a diff and you get a wall of plausible-sounding findings, most of them false, and triaging them costs more than the review saved. dsh-review tackles exactly this: it does not aim to find more, but to ensure every reported finding survives an active attempt at refutation.

The method has two stages. In Find, four distinct lenses (correctness, lifecycle and concurrency, API-contract conformance, security) each dispatch a child agent to review in parallel, and every finding must carry a concrete failure scenario rather than a style opinion. In Verify, each finding is handed to a dedicated verifier whose sole job is to refute it: read the real code, attempt reproduction, and default to "not real" when evidence is ambiguous. Only survivors are reported. On the author's two sibling plugins, a single run consumed 73 agents and yielded 49 confirmed findings against 14 that were successfully refuted; two of the confirmed ones were established only after a verifier wrote a script to reproduce the failure.

It suits teams and independent developers who demand high pre-release quality and already work inside DeepSeek Harness multi-agent workflows. This is not a per-commit lint but a comparatively expensive pre-release audit, best budgeted in minutes rather than seconds. A quick depth at roughly one-third the cost is available for small changes, and the bundled adversarial-review skill teaches the agent to judge when a full review is worth its cost and how to act on confirmed versus refuted findings differently.