xiaobright/dsh-anchored-standard

Two-phase DeepSeek Harness preset: Minimal-aligned bootstrap, then full Standard tools (Project2 98/99)

What it solves

DeepSeek V4 Pro models can exhibit different reasoning trajectories (styles of thinking) based on the tools and context provided in the initial API request. The "Minimal" tool set often leads to higher performance on complex tasks (like Project2), but using it permanently prevents the model from accessing the broader "Standard" tool set. This project provides presets that "anchor" the model's initial trajectory to the high-performance Minimal state and then "promote" the session to a larger tool set once the session is established.

How it works

The system uses a two-phase approach to manage the model's session:

  1. Bootstrap Phase: On the first request, the system provides only a minimal pair of tools (bash and str_replace_editor) and suppresses all automatic context injections (such as skill catalogs or agent instructions). This forces the model into the "Minimal" trajectory.
  2. Promotion: Once a specific signal is detected—such as the first tool call or the first assistant message—the session is "promoted."
  3. Resident Phase: After promotion, the system restores standard context injections and expands the tool catalog to include discovery tools (like dev_tool_search), allowing the model to unlock heavier tools (like web_search or subagent) on demand.

Who it’s for

Developers and researchers using DeepSeek V4 Pro (and potentially successor models) who want to maximize reasoning performance by controlling the initial model trajectory without sacrificing access to a full suite of tools.

Highlights

  • Trajectory Anchoring: Uses tool schemas and context suppression to steer the model's first reasoning chain.
  • Multiple Modes: Includes various anchoring strategies, including "Zero-Anchored" (zero tools on first turn) and "Whoami" (using a specific self-introduction turn).
  • Context Gate: A unified injection control plugin that manages when context is provided to the model.
  • Durable State: Phase state is derived from session event logs, ensuring that resume and reload operations preserve the promotion status.
  • Prefab Templates: Provides seeded rolled history to instantiate sessions in a pre-promoted state.

Related

  • Project
  • Dispatch
  • Dispatch
  • Dispatch
  • Dispatch