Introduction¶
In the streaming inference pipeline of DeepSeek Harness (DSH), incremental output fragments (deltas) from the model can sometimes be malformed or incomplete. These malformed tool-call fragments are passed directly to the block assembler, causing downstream parsing failures and interrupting the agent’s execution flow. udsy19/dsh-toolcall-stream-repair aims to address this data quality issue.
Plugin Positioning¶
This is a DeepSeek Harness plugin maintained by developer udsy19. It is positioned as a streaming data preprocessing component, responsible for repairing malformed tool-call fragments before they enter the block assembler.
Core Features¶
According to the project description, the core capability of this plugin is to intercept and repair abnormal deltas in the stream. It ensures that only expected data fragments proceed to the subsequent assembly stage.
Installation and Enablement¶
The official documentation does not provide specific installation commands. It is recommended to obtain installation instructions through the following resource:
GitHub repository
Typical Usage¶
The documentation does not provide a concrete reproducible code example. Typically, this kind of plugin exists as a stage in the DSH pipeline. Once integrated, it takes effect automatically during streaming output without manual intervention.
Applicable Scenarios and Considerations¶
- Applicable scenarios: When streaming output with a DeepSeek model frequently causes tool-call parsing failures or execution interruptions, you can consider using this plugin for data cleaning.
- Considerations: The plugin runs with the current permissions of the DSH process. Before installation, be sure to review the source code for security and confirm the license information.
Conclusion¶
By adding data repair logic to the streaming pipeline, this plugin helps improve the stability of tool calls for DeepSeek models. For more details, visit its catalog page or GitHub repository: https://www.skillhub.cn/plugins/udsy19/dsh-toolcall-stream-repair.