AI Agent Hub
Back to plugins
🧩

dsh-evolver

admin-security Updated 2026.09.04

Run the following command in DeepSeek Harness:

dsh plugin install cofy-x/dsh-evolver

Paste the following prompt into your AI chat to install this plugin:

Run dsh plugin install cofy-x/dsh-evolver in your DeepSeek Harness instance to install this plugin from https://github.com/cofy-x/dsh-evolver, then restart the target profile to activate it.

About this plugin

In DeepSeek Harness agent workflows, tool-call failures are the norm, yet the experience they carry is usually lost the moment a Session ends. dsh-evolver turns those bounded failure facts into reusable, auditable strategy guidance that an operator reviews and promotes, without ever letting the model autonomously edit source code or execute commands.

The lifecycle is deliberately explicit: a deterministic offline provider extracts bounded facts from failed results, canonicalizes them under a versioned SHA-256 signature, and aggregates repeats into patterns. Once a pattern reaches the configured occurrence threshold, a reviewable strategy proposal is generated, passes a deterministic safety verifier, requires human acceptance, and is only then promoted into future Sessions through DSH's logged injection path. Every step is reversible; a full audit trail lives in a local JSONL file, and nothing ever leaves the machine.

This is aimed at operations engineers and platform architects running multi-agent DeepSeek Harness environments who want agents to learn safely from repeated failures, without introducing the nondeterminism, external-data exposure, or cost that come with embeddings or LLM-based classification.

Use Cases

  • Distill recurring tool-call failures into reusable strategy guidance across multi-agent DSH profiles
  • Promote and roll back strategy guidance in production without losing the audit trail
  • Accumulate agent experience deterministically, with no external LLM or embedding dependency

Best For

  • Ops engineers running multi-agent DeepSeek Harness platforms
  • Platform architects seeking deterministic, offline, auditable agent learning
  • Security and compliance teams requiring full audit trails with no autonomous code changes