What it is
Harness loop where Jev, not the LLM, is the scheduler: before each step Jev reads the same session and tool history and chooses an existing Action, asks the main model, or ends the turn.
What it does: Reusable Actions (commands / scripts with declared purpose and verification) for start, test-and-build, CI repro, local release; the main model only handles what no Action covers.
Why it belongs in the showcase: A supervisor-shaped agent loop with a bilingual README and skill-style Action reuse on a non-Claude harness.
Caveats: Author says repeat work can run with zero main-model calls — author-reported; DSH + Node 24 specific.
Evidence
Public MIT README (Simplified Chinese primary + README.en.md); Jev decides each step, the main model handles novel work, and solved paths are saved as Actions for later sessions.
Caveats
- Targets DeepSeek Harness (DSH 0.2.0-rc.1) and Node.js 24.14+ per README
- Author says repeat tasks can finish with zero main-model calls — author-reported workflow runs, not benchmarked here
- Action library and thresholds are author-tuned; verify Actions before reuse
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.