dsh-agent-evolution

In multi-Agent deployments, teams often need to compare how different preset combinations of prompts, tool sets, skills, and compaction strategies perform on the same standalone task. Existing mechanisms either lack auditable execution evidence or jump straight into an automatic scoring-and-selection loop whose permission boundaries are hard to govern. dsh-agent-evolution breaks experimentation down into minimal primitives: it only supports running a task with one preset and running a baseline-versus-candidate pair sequentially, returning two records without scoring, selecting a winner, editing presets, or publishing changes. Deployments decide when and where to introduce subsequent evaluation layers.

Three runtime tools cover discovery, single execution, and sequential comparison. Every run returns the preset id, child session id, stop reason, elapsed time, final assistant content, and a persistence acknowledgment flag; child logs are flushed before the Agent handle is disposed. In comparison mode both presets are validated and authorized before either child starts, the candidate runs regardless of the baseline stop reason, and any creation, persistence, or cleanup exception halts the comparison immediately. Experiment children carry standard lineage metadata (parentSession, delegationDepth, agentPreset) yet do not appear in the SubAgent catalog or control APIs, avoiding semantic conflict with the existing subagent system.

On the configuration side, allowedPresets accepts either the system-trust roster or an explicit allowlist whose ids are all resolved at plugin load time, with unknown or broken entries blocking activation. Parameters such as maxDepth and maxTokens are rejected during configuration or request validation when invalid. This plugin is well suited for platform operators building incremental Agent governance, engineering teams wanting A/B comparisons inside independently scoped permission domains, and architects planning to introduce evaluators, failure attribution, candidate authoring, and promotion or rollback as separate, independently permissioned layers.