dsh-evolve

The DSH Agent Loop often hits the same wall repeatedly during multi-turn tasks: the same tool calls keep failing, the same no-progress trajectories recur, and user corrections are forgotten by the next run. dsh-evolve sits alongside the Agent Loop, using deterministic signals to detect repeated errors, duplicate tool invocations, and stuck trajectories, turning every repeated failure into a locally retrievable experience record without training a model or modifying DSH Core.

Its capabilities come in three layers. First, the reset-v1 strategy reset: when a clearly stuck trajectory is detected, a generic strategy reset is injected once, subject to cooldown and global caps, or the plugin can operate in observe-only mode with no injection. Second, the background experience sidecar: after a session ends, it asynchronously mines user corrections, successful procedures, and failure patterns while the main Agent path continues without waiting. Third, bounded retrieval: results are filtered by relevance, confidence, and compatibility, returning at most three temporary summaries that are discarded once the task finishes. A Skill is only created or activated through explicit distillation and user confirmation; ordinary Experience never alters future behaviour on its own.

It is built for DSH power users who want their Agent to learn from repeated failures locally, with zero cloud uploads, zero credential reads, and zero third-party dependencies. The security model is explicit: it does not disable the Sandbox, bypass Approval, execute community code, or auto-activate remote content. Observe first, act only when you confirm, always.