DSH Session Lab

Evidence packaging, Skill evaluation, and controlled trajectory comparison for DeepSeek Harness sessions.

English | 中文

Turn a successful DeepSeek Harness session into reusable evidence, measurable Skills, and controlled comparisons.

dsh-session-lab is an independent, pre-1.0 community bundle. It is not an official DeepSeek product and is not endorsed by DeepSeek. It registers three complementary workflow Skills in the DSH runtime catalog:

When you need to…	Use	You get
Share or archive one run	dsh-capsule	A verified, privacy-aware .dshc evidence bundle
Turn a successful run into a reusable workflow	dsh-teach	A candidate Skill plus independent baseline/treatment evaluation
Understand why two runs differ	dsh-time-machine	A controlled trajectory comparison with causal limits stated

The helpers use Python 3.10+ and the standard library only. They do not add a Host session-import endpoint, execute capsule contents, or claim exact model replay.

30-second demo

Install the bundle, restart the DSH profile, then ask for one concrete workflow:

▼
bash
复制代码
dsh plugin --profile web add github:zhangguiping-xydt/dsh-session-lab

▼
text
复制代码
Use $dsh-capsule to package and verify this DSH export.

The result is a .dshc archive containing a manifest, redacted session events, integrity hashes, and (when explicitly selected) a workspace patch. For a successful workflow, try:

▼
text
复制代码
Use $dsh-teach to extract and independently evaluate a Skill from this successful session.

For a controlled comparison, start two DSH branches from the same completed turn and ask:

▼
text
复制代码
Use $dsh-time-machine to compare these two sessions from their common completed turn.

Each Skill explains its required input, output, safety checks, and known limits. Read the relevant SKILL.md before handling real session exports.

Requirements: Node.js 22.19+ (or 24+) for DSH/plugin installation and Python 3.10+ for the Skill helper scripts.

Install

Install directly from GitHub into an existing DSH profile:

▼
bash
复制代码
dsh plugin --profile web add github:zhangguiping-xydt/dsh-session-lab

For a reproducible install, pin the reviewed release tag:

▼
bash
复制代码
dsh plugin --profile web add github:zhangguiping-xydt/dsh-session-lab#v0.1.0

If the dsh command is not installed globally, run the latest published CLI through npx:

▼
bash
复制代码
npx --yes @deepseek-ai/dsh@latest plugin --profile web add github:zhangguiping-xydt/dsh-session-lab

The package is currently distributed through GitHub. npm publication is not required; once an npm release exists, the shorter dsh plugin --profile web add dsh-session-lab form will also work.

For local development from a source checkout:

▼
bash
复制代码
dsh plugin --profile web add /absolute/path/to/dsh-session-lab
dsh plugin --profile headless add /absolute/path/to/dsh-session-lab

If an npm release is published later, the GitHub spec can be replaced with dsh-session-lab. Restart an already running profile after installation. Verify the composed layer before use:

▼
bash
复制代码
dsh --profile web --dump-config | rg '# == dsh-session-lab'

The bundle uses the host ctx.skills registry, so it does not depend on project-directory watchers. A source-only alternative is to copy an individual directory under <project>/.dsh/skills/ or ~/.dsh/skills/; restart DSH if an existing session catalog does not refresh.

After installation, the bundle appears as a mounted and enabled session-lab entry in Settings → Plugins → Plugin list:

What the reports look like

The public evaluation fixture for dsh-teach contains six paired tasks and twelve fresh DSH sessions:

▼
text
复制代码
baseline passed:   1/6
treatment passed:  6/6
safety failures:   0
false positives:   0
routing misses:    0

This is an example evaluation, not a guarantee for every project or model. Re-run the protocol on your own tasks before promoting a generated Skill. See the full report, Chinese case study, and evaluation protocol.

Project map

This repository is one installable DSH bundle containing three independently documented Skills:

▼
text
复制代码
dsh-session-lab/
├── dsh-capsule/       # package and verify portable session evidence
├── dsh-teach/         # extract and evaluate reusable Skills
└── dsh-time-machine/  # compare controlled session trajectories

Use

Invoke a Skill explicitly or describe a matching task:

▼
text
复制代码
Use $dsh-capsule to package and verify this DSH export.
Use $dsh-teach to extract and independently evaluate a Skill from this successful session.
Use $dsh-time-machine to compare these two sessions from their common completed turn.

Read the selected SKILL.md before execution. Session exports and generated reports may contain sensitive content even after pattern-based redaction.

Verification

▼
bash
复制代码
python3 -m pip install -r requirements-dev.txt
coverage run --branch -m pytest
coverage report --fail-under=60
ruff check .
ruff format --check .
npm test
npm pack --dry-run

pyproject.toml configures pytest and Ruff only. The project does not publish a Python wheel; the standalone Python helpers are shipped inside the npm/DSH bundle.

The suite contains synthetic security and archive fixtures plus a checked-in real-model evaluation. The full dsh-teach example records 12 fresh DSH sessions across six paired tasks: baseline passed 1/6, treatment passed 6/6, with zero safety failures, false positives, or routing misses. See the evaluation report.

Build a release artifact

▼
bash
复制代码
mkdir -p dist
npm pack --pack-destination dist
dsh plugin --profile headless add ./dist/dsh-session-lab-0.1.0.tgz

The CI workflow validates Python 3.10–3.13, package cleanliness, Skill structure, checked-in evaluation evidence, and installation into the latest published DSH profile.

Security

Treat raw exports, patches, images, and reports as sensitive. Automatic replacement is not anonymization, and SHA-256 integrity does not authenticate a publisher. See SECURITY.md before sharing artifacts.

The repository and each standalone Skill directory are licensed under MIT; per-Skill licenses allow independent copying.

See CONTRIBUTING.md and CODE_OF_CONDUCT.md before participating, CHANGELOG.md for release history, and RELEASING.md for the public-release checklist.