dsh-self-evolution

The hardest part of evolving an Agent Profile is rarely “what to change” — it is “how to roll back when a change makes things worse.” dsh-self-evolution builds an iteration loop on a frozen Benchmark: each round the Optimizer proposes candidate file edits, the Evaluator scores them in a fully isolated context, and a candidate is accepted only if its score is strictly greater than the current baseline (using >, not >=); otherwise the system rolls back to the last verified snapshot. Every decision that matters — frozen digests, the Case×Run evaluation matrix, monotonic versioning, and snapshot validation — runs in deterministic host code, never through an LLM, eliminating the risk of a model convincing itself it improved. The plugin enforces context isolation across the three roles: private rubrics never leak into the Target or Optimizer prompts, inherited tools are cleared to { allow: [] } for Evaluator and Optimizer, and every candidate write passes through a whitelist, path-traversal guard, and a per-Profile mutex lock, making each change an auditable, reversible transaction. Built for teams and solo developers already running agents on DeepSeek Harness / Cordis who have a stable evaluation set and want to turn prompt tuning from manual trial-and-error into a repeatable, safe engineering workflow.