dsh-skill-evolution

Every time an agent finishes a complex task, works around a failure, or repeats the same tool chain, that knowledge evaporates with the session and the next identical challenge starts from zero. dsh-skill-evolution adds a self-evolving skill layer to DeepSeek Harness: at the end of each successful turn it quietly inspects the execution trace, and when any of three signals (complex, recovered, repeated) fires, an LLM reviewer decides whether the episode is worth crystallizing into a reusable skill.

The real strength goes well beyond a one-shot save. A strict review gate and a hard do-not-capture list keep junk out. When new evidence lands on an existing skill, the reviewer merges the fresh lessons into the original entry and bumps the version instead of creating a duplicate; semantic deduplication (FOLD_INTO) catches near-identical lessons reached through entirely different tool sequences. In the Settings panel every threshold sits inside a single visual pipeline diagram, and review prompts can be edited in place with a one-click reset to built-ins.

Built for developers and teams who want their agents to get smarter with use. No manual skill documentation required. Let the agent do its job and the plugin quietly distills, iterates, de-duplicates, and grows a living, versioned skill catalog in the background.