AI Agent Hub
Back to plugins
🤖

dsh-gpt-file-discipline

Model Inference Updated 2026.08.22

Run the following command in DeepSeek Harness:

dsh plugin install youxi-huang/dsh-gpt-file-discipline

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

Run dsh plugin install youxi-huang/dsh-gpt-file-discipline inside DeepSeek Harness to install the plugin; the source repository is https://github.com/youxi-huang/dsh-gpt-file-discipline . After installation, mount it into your target DSH profile as a Cordis configuration patch plugin row.

About this plugin

When you route through both GPT and DeepSeek in a single DSH pipeline, a recurring pain point is that each model family needs a very different slice of system-prompt discipline. GPT models tend to browse unrelated files speculatively and benefit from an explicit file-discipline nudge; DeepSeek models occasionally drift in their internal thought language and need a lightweight anchoring line. Hand-coding conditional branches per route is tedious and easy to get wrong.

The plugin solves this by listening downstream of the waterfall so it always sees the provider and model actually chosen for the current step, not an earlier candidate. On a GPT match it appends one file-discipline paragraph; on a DeepSeek match it appends a thought-language paragraph; on any other route it returns the assembly completely untouched. No static section is ever registered, so an unmatched route incurs zero extra tokens, sections, or filtering overhead.

It is built for developers who mix multiple model providers in a single DSH profile, want per-route prompt tweaks without a build step, and prefer a single-file MIT-licensed entry point that drops in with one plugin row and is removed just as cleanly.

Use Cases

  • Auto-inject route-specific prompt sections when mixing GPT and DeepSeek in one DSH pipeline
  • Attach discipline constraints to a specific model family without altering prompt assembly on other routes
  • Apply lightweight prompt enrichment via a single plugin row with no build or prepare step

Best For

  • Inference developers running multi-provider pipelines inside a single DSH profile
  • Engineers who want per-route prompt tweaks without writing conditional branching logic
  • Users who prefer zero token overhead and plug-in-and-unplug-in simplicity over heavyweight frameworks