dshX

DeepSeek V4 Pro performance is tightly coupled to the tool catalog visible in the first API request. Official benchmarks show the Minimal preset (exact RL prompt and schemas) scoring 99/96 on Project2, while the Standard preset with roughly 25 tools drops to 91. Teams that stay on Standard risk alignment drift as the model is pulled toward standard-like behavior by the large tool surface; switching to Minimal, they lose the ability to extend tools on demand. dshX addresses this trade-off with an "anchor first, promote later" layered strategy that keeps the initial request under the strictest alignment constraints while preserving tool richness for the rest of the session.

The core mechanism lives in the built-in prefab-anchored-standard preset. On session creation, dshX pre-fills the blank conversation with a verified anchored trajectory so the user's first real task picks up from a known-good trace rather than starting cold. The first request exposes only the Minimal tool pair (persistent bash plus str_replace_editor) at the adapter's default output budget, with all auto-injected context-stripped: skill catalogs, AGENTS.md, time/tmux, hooks, and the host skill broadcast are all gated out. Once a promotion signal fires, the directory narrows to the Minimal pair plus three discovery tools, keeping heavyweight capabilities exactly one dev_tool_search away instead of dumping all 25 tools at once. Subagents follow the same policy, compaction returns the session to a controlled phase, and the lifecycle stays in a "small and stable" tool window throughout. dshX also ships the J-Space Cognition Suite built-in, available on demand via skill_search and skill_load in prefab sessions rather than auto-broadcast. Every change is made at the distribution layer with zero modifications to the deepseek-harness core. Measured on Project2 V4.1b (reasoningEffort=max), consecutive runs score 98/99.

Ideal for developers who use DeepSeek V4 Pro extensively for terminal-based coding and have observed standard-like drift in benchmark results, and who want to lock in first-request alignment quality without giving up on-demand tool extensibility. Users migrating from dsh-TUI can switch seamlessly without relocating session data or preferences.