AI Agent Hub
Back to plugins
🤖

dsh-session-slm-router

Model Inference Updated 2026.09.03

Run the following command in DeepSeek Harness:

dsh plugin install NinjaSln-labs/dsh-session-slm-router

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

Run dsh plugin install NinjaSln-labs/dsh-session-slm-router in your DeepSeek Harness (cordis) terminal to install; the source lives at https://github.com/NinjaSln-labs/dsh-session-slm-router .

About this plugin

In DeepSeek Harness multi-turn sessions, the model tier is locked once set, so lightweight prompts and complex reasoning share the same compute budget. dsh-session-slm-router addresses this by invoking an external vertical-small-model CLI every turn to classify utterance strength, building a real-traffic data foundation for future dynamic routing.

The plugin ships in shadow mode by default: it only predicts and appends to a local JSONL shadow log while leaving the active model completely untouched. Log entries carry the predicted tier, confidence, decision label, and PII-masking summaries (SHA-256 prefix plus an 80-char preview), giving teams an offline review trail. When you are ready to take the risk, switch to weak-only mode to enable one-directional downgrade switching; the plugin will never upgrade or alter the user’s persisted model preference. A built-in slot health cache re-orders candidates based on provider liveness and live availability probes, so requests never land on dead endpoints.

It is built for developers and platform operators already running the cordis host with a deployed vertical-small-model CLI who want a observe, grayscale, and roll-out cadence with zero disruption. The plugin has no extra runtime dependencies beyond the host and the CLI, and its test suite uses the Node built-in runner with no network calls, keeping integration overhead minimal.

Use Cases

  • Observe which lightweight turns in multi-turn sessions could be offloaded to a small model
  • Grayscale the downgrade path and measure latency and cost deltas on the user side
  • Replay shadow logs offline to audit CLI prediction accuracy and slot health cache behaviour

Best For

  • Developers who have already deployed the vertical-small-model CLI on the cordis host
  • Platform engineers who want to introduce SLM routing risk-free with an observe-then-rollout cadence
  • Algorithm teams that need per-turn anonymized routing decision logs for offline evaluation