dsh-skill-forge

Every time a complex conversation ends in DeepSeek Harness, the reasoning paths, debug tips, and best practices gained along the way vanish with the session. dsh-skill-forge tackles exactly this gap: it systematically distills one-off conversational experience into structured, verifiable, traceable, and reusable Agent Skills, so the next time a similar task appears you can recall a proven skill instead of starting from scratch.

Six specialized agents work in a coordinated pipeline. The Extractor pulls patterns out of conversations, the Generator drafts skill documents, the Verifier runs multi-dimensional validation, the Refiner performs targeted rewrites based on failure feedback rather than full regeneration, the Adversarial Test Generator crafts challenging cases for co-evolution, and the Auditor handles security and duplicate checks. Every skill must pass seven quality gates before activation. On top of that, three evolutionary engines keep raising the ceiling: Darwin hill-climbing for single-skill optimization, Taotie five-stage cross-skill fusion, and CoEvo alternating skill-versus-testsuite evolution. A Dreaming engine performs off-peak batch optimization and health checks during idle time. The smart recall engine dynamically selects the most relevant skills per conversation turn under a token budget, while four-layer security defense and full provenance tracking keep every skill auditable and traceable.

It is built for heavy DSH users who repeatedly complete engineering, writing, or analysis tasks and want to grow a personal skill library over time. It also suits tech leads who need to turn tacit knowledge scattered across chat logs into shared, versioned team assets.