Back to directory

dsh-anchored-standard

A frozen, community-built set of DeepSeek Harness agent presets that anchored the DeepSeek V4 Pro model's tool-use behavior on a minimal starting condition before promoting it to a fuller tool catalog.

dsh-anchored-standard

dsh-anchored-standard is a set of experimental agent presets for DeepSeek Harness that anchor a session's model behavior on a minimal starting condition, using a real minimal tool schema with no automatically injected context, then promote the session to a small resident tool catalog once it proves durable, unlocking heavier standard tools on demand. It includes a base mode, two live-anchor variants and a seeded prefab mode.

It is a community project, not an official DeepSeek release, and is not affiliated with or endorsed by DeepSeek. The project is frozen as of September 2026: after DeepSeek retired the V4 Pro model these presets were tuned for in favor of V4.1 Flash, the maintainers judged that the successor model did not need the same patching, based on earlier dose-response probes, and stopped accepting new issues or pull requests, while keeping the repository's history and data online.

Its model-agnostic components, such as a unified context-gate injection control and a prefab pipeline, remain documented as still valid.

More in Agent skills

See all 581
  • Provenance gate

    github.com

    A deterministic authorization gateway that sits between an AI agent and its tools, tracing where each parameter of a side-effect call came from, and blocking the call or asking a human when the source does not check out.

  • A markdown framework of system rules that restates Isaac Asimov's Four Laws of Robotics as a set of prioritized loss calculations for large-language-model agents to follow, published as a discovery file for agents and crawlers to read.

  • Vitruvius

    github.com

    An engineering research agent that runs a discover, read, synthesize, verify and review loop over technical questions, reading sources directly rather than inferring from titles, and publishing its own scored verifier benchmark, weaknesses included.