AI Agent Hub
Back to plugins
dsh-learning-bundle preview

dsh-learning-bundle

Memory Updated 2026.08.24

Run the following command in DeepSeek Harness:

dsh plugin install B1lli/dsh-learning-bundle

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

Run dsh plugin install B1lli/dsh-learning-bundle inside DeepSeek Harness to install this plugin; the source code is available at https://github.com/B1lli/dsh-learning-bundle .

About this plugin

Repeatedly correcting a coding agent is painful not because of any single mistake, but because corrections do not survive safely into the next session. A global, invisible memory can leak one project convention into another; an opaque injection makes the next error nearly impossible to diagnose; automatic activation can turn a bad guess into a durable rule. dsh-learning-bundle addresses this with a deliberately narrow correction-delivery loop: record a correction as a candidate, adopt it explicitly, recall it only within the right profile, workspace, and session scope (failing closed), and let the current task instruction always override a learned default.

The core capability is a complete lifecycle of record, adopt, recall, and reconstruct. Every correction delivered to the model is written into the session log with a source.kind marker, so at any time you can replay exactly what the model saw and in which context. The plugin carries zero external dependencies, does not auto-extract memories, and does not pretend to be a general RAG system. It provides one inspectable, auditable reference implementation for correction delivery.

It is aimed at heavy DeepSeek Harness users, independent developers, professional engineers, and plugin or platform authors who need to validate agent behavior. If you do not need a full memory control plane and only need a verifiable delivery evidence chain in repeated-correction scenarios, this plugin offers a narrow and transparent starting point.

Screenshots

Use Cases

  • After repeatedly correcting a package-manager preference, keep the fix scoped to one workspace instead of a silent global override
  • Reconstruct from the session log exactly what correction the model received and in which context
  • Verify that a current explicit instruction always wins over an adopted learned default

Best For

  • Heavy DeepSeek Harness users who correct coding-agent behavior on a daily basis
  • Plugin and platform authors who need to validate agent delivery behavior
  • Professional engineers collaborating with coding agents who value auditable correction trails