dsh-session-lab

A DeepSeek Harness session ends, and the results scatter: no shareable archive, no mechanism to distill what worked into a repeatable workflow, and no way to explain why two runs diverged. dsh-session-lab fills those three gaps with independently invocable Skills registered in the DSH runtime catalog, so you can trigger the right workflow simply by describing the task in natural language.

The bundle centers on three complementary Skills. dsh-capsule packages a single run into a .dshc evidence bundle with integrity hashes and redacted events, ready for audit or sharing. dsh-teach distills a successful run into a candidate Skill and automatically runs an independent baseline-versus-treatment evaluation, producing quantifiable improvement data. dsh-time-machine forks two DSH branches from the same completed turn to perform a controlled trajectory comparison, while explicitly stating the limits of causal claims. All three rely solely on the Python 3.10+ standard library, add no extra services, do not execute bundle contents, and keep the attack surface minimal.

This bundle is well suited to teams that want to turn DSH sessions into reusable process assets, reviewers who need a concrete evidence trail to justify decisions, and researchers running controlled comparisons of model behavior where confounding variables must be isolated. Each Skill documents its required inputs, output format, safety checks, and known limitations in its SKILL.md; reading the relevant document before handling a real session export is strongly recommended.