xiaobright/dsh-anchored-standard

This repository contains experimental presets for DeepSeek, an AI model. The project is community-built and not officially affiliated with DeepSeek. The core idea is to control how the AI model behaves during its first interaction in a session, then gradually unlock more capabilities as the session becomes stable. The problem the project addresses is that DeepSeek V4 Pro reacts strongly to which tools it can see. In testing, a minimal tool setup produced better scores (98 and 99) than the standard tool setup (91 and 92). However, staying on the minimal setup permanently means losing access to the broader tools available in the standard setup. This project splits the difference: start the session with the minimal tool configuration to get the better initial behavior, then switch to the standard tool set once the session is established. The repository offers three modes. The base mode, called Anchored Standard, starts with two tools visible to the model. A second mode, Zero-Anchored Standard, starts with zero tools and uses one fixed initial turn, costing one extra model call. A third mode, Whoami Standard, starts with zero tools and asks the model to introduce itself, also costing one extra model call. Each mode directory is self-contained and can be installed on its own. The mechanism works in two phases. The first request uses the minimal tool pair and strips away auto-injected context like workspace digests and skill reminders. This influences the model's initial reasoning style, producing "We need" opening lines instead of "Let me" lines. Once the session records a durable event, meaning the model makes a tool call or sends a reply, the preset promotes to a resident catalog. This catalog includes the initial tools plus discovery tools that let the model unlock heavier capabilities on demand. The phase state is derived from durable events, so resuming or reloading a session preserves it. The results section reports two test runs on a specific benchmark, scoring 98 and 99. The author notes this is reproducible evidence for the specific task, not a claim that the approach universally improves performance across all models or workloads. Full methodology lives in a separate repository called modeltest. The full README is longer than what was shown.