Introduction¶

In DeepSeek Harness (DSH), a failed tool call often means a full model retry: the error is returned, the model rereads the entire conversation context, and then tries again. In large workspaces, this type of retry resubmits about 180k input tokens. A failure rate of around 5% can increase token costs by more than 2x. The merenguesl/dsh-tool-normalizer plugin sits on the tools/execute pipeline and performs fixes before the call reaches the model: it fills in missing description values, executes forgotten file reads and retries edits, resolves relative paths, and appends actionable hints.

Plugin Overview¶

This plugin acts as a self-healing layer for model tool calls, fixing failures before they reach the model.

• Maintainer: merenguesL
• License: MIT
• Where it works: tools/execute pipeline
• Additional feature: includes a Web UI diagnostics dashboard

Core Features¶

The plugin provides the following capabilities:

1. Schema Auto-Healing
   Wraps commands into run_code, fills in missing description values, and removes Markdown fences.

2. Code-Mode Direct Tool Bridging
   Reschedules UNKNOWN_TOOL calls.

3. Range & Path Normalizer
   Resolves relative paths and checks boundaries.

4. Real-Time Telemetry & Statistics Tracker
   Real-time telemetry and statistics tracking.

Usage Examples¶

• Fixes schema mismatches, such as converting command to code, or filling in missing description values.
• Retries safe file mutations and bounded range retries.
• View KPI cards, rankings, and traces in settings.

Installation and Enabling¶

Install via GitHub:
https://github.com/merenguesL/dsh-tool-normalizer

Applicable Scenarios and Precautions¶

• Suitable for DSH developers who need high reliability.
• The plugin runs under the current DSH process permissions.
• Check the source code and license before installing.

Conclusion¶

This plugin reduces retry rounds and token costs by automatically fixing tool call errors, improving agent stability. See GitHub for details.