The philosophy behind DSH is “everything is a plugin.” When developing with local large language models, models often make repeated mistakes (goldfish memory) or waste compute in infinite loops. dsh-host-rsi is a host-side plugin designed to address this pain point, enabling local models to accumulate experience during a session and automatically correct their behavior.

Plugin Positioning

This plugin belongs to the model inference category and is maintained by the developer pk14742952-AD. It is a pure host plugin that does not touch model weights. Its core value is to help local large language models become “stronger with use” within a session and prevent them from falling into infinite loops.

Core Capabilities

The plugin mainly provides the following capabilities:

  • Self-evolution: Through an L1/L2/L3 signal staircase reliability-weighting mechanism, it hierarchically processes the model’s self-inspection conclusions, tool successes and failures, user corrections, and long-term preferences, distilling high-value experience.
  • Convergence guidance: When detecting infinite loops, it automatically injects convergence prompts, preventing the model from repeatedly attempting already verified tasks and saving compute resources.
  • Five-state delivery guardrails: It intercepts five types of delivery failure states (EMPTY, TRUNCATED, FIXED_POINT, SAME_PLAN, and SHELL_LOOP), ensuring outputs are complete and valid.
  • Persistent memory: It uses in-memory cache reads plus background debounced writes to ensure the hot path is never blocked, while also providing crash self-recovery capabilities.
  • Project-aware instruction deduplication: It groups by project/workspace to prevent the memory store from bloating due to duplicate instructions.
  • Zero intrusion: Running as a host plugin, it does not interfere with the model’s inference process.

Installation and Activation

At present, the plugin does not provide a CLI installation command and must be installed via the GUI:

  1. In the DSH interface, go to Settings → Plugins.
  2. Install dsh-host-rsi.
  3. Restart DSH.
  4. Go to Settings → Plugins → DSH HOST RSI Memory → click Open Dashboard.

Typical Usage

After installation and activation, you can verify and use it in the following ways:

  • View the dashboard: After restarting DSH, open the dashboard on the plugin settings page to view captures, injections, interceptions, and recent trigger events in real time.
  • Run the standalone demo: In the project root directory, run the following command to start the standalone demo service:
    node demo-serve.mjs
  • Call the diagnostics API: Call rsi_events to view trigger diagnostics.

Notes

  • Compatibility: General-purpose mode, compatible with both API-based and local models, and tuned by default for local 27B/Flash models.
  • Data source filtering: It captures only messages where source.kind === 'user', preventing system prompts from polluting memory.
  • Fallback strategy: If the plugin directory is not writable, it automatically falls back to pure in-memory mode without affecting functionality.
  • Performance optimization: It uses background debounced writes, ensuring that the hot path (pre-step / turn-end) is never blocked.

Conclusion

Through its signal staircase and delivery guardrail mechanisms, this plugin provides a memory layer and convergence layer for local models in DSH. For more detailed principles or the changelog, refer to its GitHub repository.