Introduction

DeepSeek Harness (DSH) adopts a “everything is a plugin” architecture. When building agents, strategies often need to evolve in response to runtime feedback. How to manage this evolution and ensure that changes remain auditable and controlled is a common development challenge. The dsh-evolver plugin uses a verification gating mechanism to convert bounded execution failures into reviewable strategy proposals.

Core Features

This plugin mainly addresses the control and tracking of strategy evolution:

  1. Audit and verification gating: Provides auditable self-evolution capabilities, requiring manual acceptance and promotion before guidance enters subsequent Sessions.
  2. Pattern classification: Classifies failed context summaries into precise signature patterns.
  3. Proposal management: Generates proposals and stores bounded diagnostic facts and Session references.
  4. Context control: Only promoted text enters the model context through the public, logged agent.inject() channel.

Installation and Enablement

Add the plugin to the Web or Headless profile in DSH:

dsh plugin --profile web add dsh-evolver@alpha

After installation, it is recommended to run a configuration check:

dsh --profile web --dump-config

Typical Usage

After installation, use the following commands through the DSH client to view, evaluate, and decide how to handle generated strategy proposals:

  • /evolve list: List all proposals.
  • /evolve patterns: View all classified failure patterns.
  • /evolve pattern <pattern-id>: View details of a specific pattern.
  • /evolve show <proposal-id>: Show the contents of a specific proposal.
  • /evolve accept <proposal-id>: Accept the proposal.
  • /evolve promote <proposal-id>: Promote the proposal so it affects subsequent Sessions.
  • /evolve evaluate <proposal-id>: Evaluate the performance of the proposal.
  • /evolve reject <proposal-id> <reason>: Reject the proposal with a reason.
  • /evolve rollback <proposal-id> <reason>: Roll back already effective guidance.

Use Cases and Considerations

This plugin is suitable for developers who need strict control over strategy change processes.

Important Notes:

  • Permissions and execution: The plugin runs with the permissions of the current DSH process. It does not automatically edit DSH, plugins, or user code, and it does not hold model credentials or tool permissions. It cannot execute proposal text or rewrite validators, evaluators, approval policies, or security policies.
  • Limitations: Its fixed diagnostic templates and structural validators do not establish improved task success. Evaluation is based on a comparison of failure rates within the exposed scope, used to support human decisions rather than to trigger automatic rollback.
  • Data security: Audit storage keeps bounded diagnostic facts and Session references, and does not store full records, reasoning, tool parameters, credentials, or complete tool outputs. Redaction is defensive and cannot guarantee that arbitrary text is free of sensitive information.
  • Concurrency: Each data directory should use one active runtime.
  • Version: The current examples use the @alpha tag. If a stable version is required, replace it with a specific version number.

Short Conclusion

dsh-evolver provides DSH with a mechanism to extract strategies from failures while keeping them under human review. By limiting changes to the text level and strictly bounding the scope of stored data, it offers a controllable path for strategy evolution.