dsh-self-harness-tools

Self-harness tooling for the DeepSeek Harness ecosystem

_{中文 · English}

Two installable DSH bundles that measure and inspect your own harness:
a probe battery for model priors, and an archive dashboard for release history.

Why

DeepSeek Harness follows an "everything is a plugin" architecture. This
repository ships two bundles built on the official Cordis
plugin model, installable with dsh plugin add and distributed on npm:

• dsh-prior-probe — frozen probe-battery tooling. It measures how much
  of an expected answer a model can produce from prior knowledge alone
  (no-harness, single-shot completions scored with deterministic
  grader-equivalent regexes).
• dsh-evidence-dashboard — an archive dashboard for your project:
  components, design decisions, rejected options, and a git-tree of the
  release history with a changelog timeline.

Features

Quick start

Install both plugins into a profile with the shortest official channel:

dsh plugin --profile demo add dsh-prior-probe
dsh plugin --profile demo add dsh-evidence-dashboard

dsh --profile demo --dump-config   # verify the "# == dsh-prior-probe" layer
dsh web --profile demo             # open Settings → DSH Evidence

Alternatives: from GitHub
(dsh plugin add github:SuTang-vain/dsh-self-harness-tools#path:plugins/prior-probe),
or from a packed tarball. See each plugin README for details.

Documentation

Evidence / 效用证据

Utility data, not promises — what the pieces demonstrably do:

• dsh-plugin-kit: 14/15 checks green across three third-party
  repositories with three different distribution shapes (npm bundle /
  official .dsh-plugin / profile snippet); the kit also runs 82 checks in
  this repository's own CI. See docs/kit-field-test.md.
• dsh-prior-probe: a deterministic offline demo produced by the shipped
  scoring code ships at
  plugins/prior-probe/data/example-score-run.json
  (5/8 scored, G=0.8, low-direction flags visible); live endpoint runs use
  prior_probe_run with your own credentials.
• dsh-evidence-dashboard: renders this repository's own release history,
  and the schema is demonstrated on a third-party project's real data —
  the six npm publishes of @deepseek-ai/dsh in
  plugins/evidence-dashboard/data/examples/versions-dsh-official.json.

Relationship to the Official Project

This repository is built on deepseek-ai/deepseek-harness
and the Cordis plugin model.

The official project provides the core agent capabilities, the plugin system,
and the Web UI. This repository primarily provides:

• Two reusable, installable bundles for the official dsh plugin flow
• Tooling to measure model priors before building a task suite
• An archive dashboard with a version tree for project history
• npm distribution with dsh.bundle / dsh.client manifests
• dsh-plugin-kit, a reusable QA kit for DSH plugin repositories

If you prefer to contribute to the Harness core itself, refer to the official
repository first.

Community

• Official DeepSeek Harness: GitHub Discussions · Discord
• This repository: report issues and open pull requests on GitHub.

License

This project is licensed under the MIT License.

Community tooling built on DeepSeek Harness. It is not an official DeepSeek
product.