SheberDavid/v4-flash-godmode-opencode-go

What it solves

This project provides a specialized agent preset for the DeepSeek V4 Flash model via the opencode-go provider. It addresses the "ghost mode" issue where the Flash model, without proper guidance, tends to produce shallow thinking, rushed execution, and low-quality output. By implementing specific guidance anchors, it transforms the model's performance into "god mode," enabling deep planning, high-quality delivery, and error-free self-testing.

How it works

The project integrates a set of optimized guidance anchors—including task classification, review, anti-digression, deep thinking, and decision loops—directly into the WEAK_FLASH persona. This static merging of guidance prevents reliance on dynamic injection mechanisms that are often unstable in newer versions of the DeepSeek Harness (dsh) framework.

Who it’s for

Developers and users of the DeepSeek Harness (dsh) who are using the deepseek-v4-flash model through the opencode-go provider and want to improve the model's reasoning depth and reliability for complex building tasks.

Highlights

• Significant Performance Boost: Increases planning depth (e.g., from 29k to 375k words in a 3D simulation task) and adds numerical verification and self-testing.
• Cross-Platform Support: Automatically adapts to Linux (bash), macOS (bash), and Windows (pwsh).
• DSH Framework Integration: Specifically fixed for dsh rc.6+ to ensure reliable persona injection.
• Targeted Optimization: Specifically tuned for Flash models to trigger deep thinking anchors while leaving non-Flash models unaffected.