dsh-missher-evolution

Every time you open DeepSeek Harness you re-state your preferences and re-correct the workflow, because the model simply does not remember last turn. Missher Evolution tackles exactly that: it distills reusable rules from your direct foreground user turns and promotes them through a three-stage verification pipeline (Candidate to Trial to Active) so that only well-confirmed patterns actually influence future instructions.

All learned state lives under a local $DSH_HOME directory, is schema-validated, and contains no raw user messages, credentials, absolute paths, or provider route names. The learning pipeline is fail-open by design, meaning if the store, classifier, or advisor hiccups a normal Harness step is never rejected. Only direct foreground user turns trigger learning; subagents, tool continuations, scheduled work, and internal plugin messages are all filtered out, keeping the boundary tight.

Built for users who work in DeepSeek Harness Desktop daily and want their personal workflow preferences to compound over time. It improves instruction and workflow reuse rather than fine-tuning or replacing the underlying DeepSeek model, and it does not import or synchronize state across agent hosts.