dsh-anchored-standard: Maintaining DeepSeek V4 Pro's 99-point performance while leveraging standard tools

DeepSeek V4 Pro's performance varies significantly depending on the list of tools exposed to the API. While the Standard preset offers a broad range of tools, its score remains at 91–92 points, whereas the Minimal preset soars to 99 points. This project is a collection of experimental agent presets designed to maintain the high performance of the Minimal preset while utilizing the rich tool capabilities of the Standard preset.

At the start of a session, only the actual Minimal tool schema is exposed, and auto-injected context is blocked to fix the model's initial reasoning trajectory. Once the first sustainable tool call or assistant message is recorded, a 'promotion' occurs, making heavier Standard tools available. This ensures both high accuracy in the initial stages and flexibility in later stages.

In addition to the default mode, various variant modes are provided, including Zero-Anchored, Whoami, Prefab, and Eternal Minimal. Each mode controls the model's behavior by adjusting the conditions of the first request (tool schema, output budget, injected reminders). Notably, the Prefab mode uses successful trajectories as seeds to start a session in an immediately promoted state without any model calls.

This project is a community-based experimental tool, not an official DeepSeek preset. Due to recent increases in API costs, active development has been halted, but the existing mechanism discoveries and tools remain valid with maintenance only. It can serve as a useful reference for developers researching performance optimization and tool management strategies for DeepSeek-based agents.