harness-self-evolution-plugin

As the DeepSeek Harness plugin ecosystem grows, deployment is only the beginning. Questions like how to quantify plugin health, when to trigger an evolution, and how to roll back safely after a failure have long lacked a systematic answer. harness-self-evolution-plugin, written in MoonBit and compiled into a stdio MCP server, runs inside the DSH runtime and turns scanning, monitoring, proposal generation, human approval, execution, and deterministic rollback into a single auditable loop, so plugin improvement no longer depends on scattered scripts or ad-hoc manual checks.

The plugin continuously tracks call latency, success rate, token usage, retry count, and user feedback. From these metrics it extracts strong signals such as three consecutive failures, a latency regression above 20 percent, or user override, as well as weaker signals like repeated parameter misuse, loop detection, and recurring preferences, then generates benchmark-driven evolution proposals. Approval is always a human step: auto_approve remains false under every configuration, making a person the sole gate against uncontrolled self-rewriting. Execution follows a strict state machine from pending through approved and executing to completed, and on validator failure a deterministic rollback restores the last verified snapshot with no half-applied residue.

If you maintain multiple DSH plugins and want a repeatable detect, propose, approve, execute, roll back workflow, or if you are a DSH plugin developer looking to reuse its sub-agent factory and signal-recognition logic, this plugin offers an engineering-grade, reversible, and auditable path for plugin health governance.