dsh-continual-evolve

The pain point in many agent workflows is not a lack of reasoning ability, but the inability to retain experience. Fixes made during a session, durable facts, and reusable procedures often disappear in the next session. dsh-continual-evolve turns such experience from session trajectories into first-class state, allowing DeepSeek Harness to refine itself within a session and reuse knowledge across sessions in a controlled way.

Its guiding principle is “the model proposes, the code guarantees.” Prompt notes, memories, skills, and subagent specifications are stored in a versioned, auditable form with schema validation, atomic writes, snapshots, and audit events. Local and global stores are separated, with promotion guards, near-duplicate rejection, and explicit approval helping keep the global knowledge base clean. Candidate refinements can be evaluated against frozen benchmark cases, and rejected changes roll back deterministically.

This plugin is a good fit for developers and teams building long-running, self-evolving agent workflows, especially when they need auditable state, reversible changes, strict admission rules, and safe injection of accumulated experience into the system prompt.