Introduction¶

In agent development for long sessions or complex workflows, tool results can easily become uncontrolled, consuming a large amount of context budget and potentially mixing in malicious instructions. dsh-jev-tools is a DeepSeek Harness plugin that uses the Jev judgment model (TypeSafe System One) as a “triage desk” to filter and judge content before it enters the context.

Plugin Introduction¶

The plugin is maintained by HorusJiang and categorized under admin-security. It does not replace the main model or generate text; instead, it classifies and processes input content through probabilistic ranking.

Core Features¶

The plugin provides the following five core capabilities:

1. Condensing Tool Output
   Condenses results from read, grep, glob, web_fetch, and web_search. If a result exceeds 2000 tokens, the plugin judges relevance paragraph by paragraph and retains relevant paragraphs. Retained passages are distributed across the result rather than being cut as a continuous segment.

2. Injection Screening
   Screens the content fetched by web_fetch / web_search and determines whether it contains instructions targeted at the AI. Once the threshold is exceeded, a reminder is attached, but the content still enters the context as-is (no blocking, no rewriting).

3. Skill Recommendation
   On the first prompt assembly of each turn, when the skill catalog has at least 15 skills, it selects at most one best-matching skill based on the latest user message (or the last three messages).

4. jev_ask
   Actively invoked by the model. Accepts any typed question and directly returns an answer with probabilities.

5. jev_gate
   Actively invoked by the model. Before declaring a task “complete,” it checks each claim for supporting evidence. This is the only place in the plugin where the fail-open policy is reversed; any ambiguous path triggers escalate.

Installation and Enabling¶

Before installation, ensure DeepSeek Harness is installed.

dsh plugin --profile web add dsh-jev-tools

Note: The --profile parameter is required. It forwards the value as-is to pnpm in that profile directory. web corresponds to the desktop / web app. Replace it with the profile you actually use. You can also install by package name on the DSH plugin page; that path completes installation and enabling in one step.

Configuration and Diagnostics¶

Once enabled, all automatic capabilities are on by default. To adjust the behavior, a TypeSafe API Key must be configured.

API Key Configuration¶

If no key is configured, the plugin remains fully dormant: it does not take effect and does not make network requests. There are three configuration methods:
1. Use the official SDK environment variable TYPESAFE_API_KEY.
2. Paste and save it in DSH settings (the key is written through the credential domain and is not echoed back).
3. Configure it via an environment variable or a .env file (resolution order: process environment > project-env > user-env > .env > managed storage).

Configuration Items¶

The main configuration items include the master switch, key variable name, judgment endpoint, model, and switches for condensing and skill recommendation. Other items, such as thresholds and allowlists, can be modified in the configuration line.

Diagnostic Command¶

Use the /jev-status command to view the enabled status, key source, judgment endpoint, judgment count, and the reason for each skip (such as task-too-vague, budget-turn, etc.).

Notes¶

1. Input Limitation: Only text input is supported; no images or audio.
2. Model Positioning: It is not another large model and does not replace the main model for writing, reasoning, or tool calls.
3. Uncalibrated Probabilities: The output probabilities are used only for ranking, are not calibrated probabilities, and should not be treated as accuracy metrics.
4. Fail-open Policy: Except for jev_gate, if other capabilities fail, they pass content through unchanged; a failed judgment does not affect the task.
5. Data Boundary: Some data (such as tool output content, fetched page content, and current task text) is sent to the configured System One endpoint (default api.typesafe.ai).
6. Key Requirement: A TypeSafe API Key is required.

Summary¶

This plugin provides output management capabilities for DeepSeek Harness based on the Jev judgment model, making it suitable for scenarios that require fine-grained control over context management, security screening, and skill routing. For more details, refer to the GitHub repository or the community catalog.