dsh-jev: Injecting Per-Turn Sentiment and Intent Verdicts into DeepSeek Harness
excalibur9527/dsh-jev
A DeepSeek Harness plugin that calls typesafe.ai systemone on every turn to detect user emotion and intent, then injects the verdict as source-tagged runtime context so the model can tune tone and priority.
dsh-jev is a DSH plugin that hands the user's latest message to the systemone detection endpoint before each turn, collects verdicts such as sentiment, urgency and routing department according to the questions you configure, then injects them as source-tagged runtime context so the model can adjust tone and priority. It fires only on steps with a new user message, so tool steps and its own injected content never re-trigger it, and identical input is reused for five minutes. A missing key, HTTP error or timeout only logs a warning and lets the conversation proceed, never blocking it; every parameter lives in the settings page alongside a live test run and a recent-calls list.
How to Install
dsh plugin --profile web add github:excalibur9527/dsh-jev- Category
- Skills & Agents
- Platform
- DSH Plugin
- Author
- excalibur9527
- Distribution
- Plugin
dsh-jev Key Features
dsh-jev Repo Summary
What Does It Do?
JEV Sentiment Analysis is a community DSH plugin for DeepSeek Harness that turns per-turn sentiment and intent detection into injectable runtime context. Before each step it sends the user's latest message to a third-party systemone endpoint, collects verdicts such as sentiment, urgency and routing department according to the questions you configure, then injects them as source-tagged context so the model can adjust tone and priority. All parameters live in the GUI settings page, and the plugin is maintained by Excalibur9527 under the MIT license.
Core Features
- Sentiment and intent detection: calls the detection endpoint once per turn, only for the user's latest message, skipping tool steps and plugin-injected context.
- Context injection: appends the verdict as source-tagged runtime context to the request so the model can adapt tone and priority.
- Configurable questions: supports yes/no, score and single-choice question types, with department, sentiment and urgency as defaults.
- Non-blocking by design: a missing key, HTTP error, timeout or network failure only logs a warning and lets the conversation proceed unchanged.
- Caching and self-trigger protection: identical input is reused for five minutes, and the plugin source tag keeps injected messages from re-triggering detection.
- Settings page and debug views: includes a live test run plus a recent-calls list showing rendered output, raw JSON, latency and the latest error.
How to Use This Plugin?
After enabling it in DSH, open the JEV sentiment analysis section in settings and enter your API key, which can also fall back to an environment variable. Adjust the endpoint, model, timeout, maximum input characters and the extra behavior hint as needed. The default questions cover department, sentiment and urgency, and can be rewritten as yes/no, score or single-choice items; turning off injection still runs detection and records results. Use the test run button to hit the endpoint with the current configuration, then check recent calls for latency and output. Note that the text of each latest user message is sent to that third-party service, so point the endpoint at your own proxy or disable detection if that matters.
This page is an independent rewrite of the plugin's official README — for authoritative documentation and the latest changes, refer to the source: excalibur9527/dsh-jev. The plugin is third-party code that runs on your machine once installed; inclusion does not imply endorsement — please review the source before installing.
DSH Plugin FAQ
