The default behavior of DeepSeek Harness Desktop may introduce uncertainty in file reading and semantic authorization. The dsh-jev-mode plugin provides an optional JEV decision mode designed to introduce deterministic file reads, bounded semantic selection, and independent authorization gating, helping developers more precisely control tool usage and permissions.
Plugin Overview¶
dsh-jev-mode is a community plugin developed by QI-seven33 (unofficial). It acts as an optional preset, combining the capabilities of standard Desktop tools and focusing on addressing the following issues:
* Deterministic file reads: When a file path is explicitly specified, the code directly controls the read operation instead of relying on the model’s semantic selection.
* Bounded semantic selection: It limits the scope of JEV’s choices in specific scenarios to avoid excessive divergence by the model.
* Independent authorization gating: It provides a separate authorization verification step for high-risk operations.
Core Features¶
The plugin mainly provides the following behavior patterns:
-
Deterministic Reads
When a task explicitly requires reading the root directory’spackage.jsonorREADME.md, the code directly selects the file and performs the read. In this case, the JEV selection logic is not invoked. -
Bounded Semantic Selection
In eligible comparative request scenarios where both fixed files are available, JEV performs one bounded semantic read selection. The read is executed only when both confidence and selection probability reach 0.9; otherwise, it falls back to normal planning. -
Authorization for Risky Operations
For high-risk operations, the system first runs hard rules. For operations with a risk level of 2+, it uses the actual human instruction history for semantic authorization verification, and the host authority’s veto has final effect. -
Native UI
The plugin integrates separate authorization and credential management interfaces, supporting configuration, saving, replacement, and removal through native modals.
Installation and Enablement¶
To install and enable the plugin in Harness Desktop, follow these steps:
- Open the Plugins page.
- Enter the installation command and enable the plugin:
github:QI-seven33/dsh-jev-mode#v0.4.2
- Fully restart Harness.
- In a new session, go to Settings -> Models -> JEV decision model -> Edit.
- Enter the TypeSafe API Key and save it.
Typical Usage¶
After enabling the plugin, different task types trigger different decision logic:
- Explicit root-file lookup: When an instruction asks to view a configuration file or documentation in the root directory, the system reads the file directly without invoking JEV semantics.
- Comparative tasks: When the request is a comparative task involving fixed files, the system performs a single semantic read only when both confidence and probability meet the required thresholds.
- High-risk operations: When executing operations with risk level 2+, the system combines instruction history for semantic authorization verification and prioritizes hard rules.
- Ordinary code or documentation tasks: For routine code writing or documentation processing tasks, the system continues to use standard planning and tools and does not automatically trigger JEV calls.
Notes¶
- Version testing: This plugin has been tested on DeepSeek Harness Desktop 0.2.0-rc.2. Other versions or operating systems may require separate verification.
- Unofficial plugin: This is a community-developed plugin. Please review the source code and license yourself.
- Uncalibrated thresholds: The authorization thresholds (0.85 / 0.55) have not yet been calibrated for the current domain, and the real-world false positive and false negative rates have not been measured.
- Credential security: Credentials provided by the environment are read-only. Removing a stored key may expose the existing
.envfallback mechanism.
Conclusion¶
dsh-jev-mode provides a more controlled decision path for DeepSeek Harness Desktop. If you need to strictly manage file read scope and authorization flows, you can try enabling this mode. For more development details and validation data, refer to the GitHub repository or the SkillHub catalog.