dsh-quality-review

When an LLM agent operates inside DeepSeek Harness, its replies can easily contain factual errors, leave parts of the user question unanswered, present contradictory reasoning, or violate explicit format and language constraints set by the user. Catching these problems after the fact is far more costly than intercepting them at the source, yet most workflows lack a lightweight, plug-and-play quality gate.

dsh-quality-review addresses this gap by hooking into the agent/turn-stopping event. It captures the assistant's latest visible reply and hands it to an independent reviewer model for audit across four independently toggleable dimensions: factual accuracy, answer completeness, logical consistency, and instruction following. If the verdict is fail, the plugin issues a concrete fix request to the agent and keeps the turn open so the model can revise its output. A hard cap of two revision rounds per turn (maxRounds defaults to 2) prevents infinite retry loops. Every configuration field ships with a safe built-in default, so the plugin works immediately with zero setup; individual keys can be overridden via cordis.patch.yml to enforce stricter or more relaxed quality standards.

If you are building user-facing or downstream-consumed agent workflows on DSH and need a consistent quality safeguard that does not rely on heavy human review, this plugin offers a lightweight quality gate that fits naturally into the existing turn lifecycle.