dsh-evolution-lab

You keep hitting the same edge cases in a DeepSeek Harness project, and turning that trial-and-error into a reusable SKILL.md still means hand-editing a file, eyeballing whether the change helps, and manually reverting when it does not. Evolution Lab automates that loop end to end: it extracts experience from project sessions, generates a quarantined SKILL.md candidate, scores baseline versus candidate inside an isolated DSH subprocess, runs held-out canary tasks, and then atomically promotes or rolls the skill back, all without a manual approval queue and without an override flag. The autonomy boundary is deliberately narrow. One Markdown Skill at a time, capped at 16 KiB, with a fixed invocation policy. The model gets a single read-only status tool and can neither vote, promote, weaken policy, or grant exceptions. On the data side, raw session logs stay DSH-owned; Evolution Lab persists only bounded, HMAC-pseudonymized envelopes that pass a second leak scan before storage or engine submission, and it never ships credentials, absolute home paths, or canary fixtures to the native model or an optional xskill bridge. This plugin fits teams that already run DSH, want project experience to keep compounding into auditable skills, and need isolated evaluation plus full rollback history. It will not fly fully autonomously, but if what you want is automatic evolution with evidence, enforced boundaries, and an immutable promotion log, V1 deliberately trades breadth for safety and ships exactly that.