timwhitez/dsh-self-evolving
Evidence-first DSH self-evolution engine (standard Cordis controller/service): bounded generation of Cordis candidate plug-ins, one-time real Loader isolation admission, Harbor evaluation, crash-recoverable journal lineage; 291 single tests + 36 Loader E2E
dsh-self-evolving is an evidence-first, crash-resumable self-evolution engine for DeepSeek Harness: it generates bounded Cordis plugin candidates, runs them through isolated real-Loader admission, evaluates them with Harbor, and preserves an auditable lineage. Every candidate is treated as untrusted — the controller, evaluator, budget, dataset split, and safety policy form a trusted computing base, and a result is accepted only when its source identity, evidence, cost, lifecycle, and recovery path reconcile. v0.2.0 verifies stable iteration and a measurable fixed-replay engineering effect (running candidates through a fixed replay), while explicitly declining to claim benchmark improvements.
Install
dsh plugin --profile headless add @dsh-self-evolving/core@0.2.3npm @dsh-self-evolving/core 0.2.3 verified 2026-09-02 (maintainer timwhitez; README EN primary with 简体中文, repo timwhitez/dsh-self-evolving; 291 unit + 36 E2E tests per audit docs). Controller installs into a headless profile per README: dsh plugin --profile headless add @dsh-self-evolving/core@0.2.3 (pnpm 11.7.0 through Corepack). Source/dev install: git clone + pnpm setup:source, then pnpm dsh-self-evolving init/doctor/run --state-dir .... v0.2.0 explicitly does NOT claim a Terminal-Bench score improvement, sealed promotion, leaderboard result, or SOTA.
Compatibility
DSH headless profile; pnpm 11.7.0 via Corepack; controller + evaluator + budget as trusted base.
Details
- Repo: timwhitez/dsh-self-evolving
- Category: Other
- Stars: 2
- Version: npm @dsh-self-evolving/core 0.2.3
- Last push: 2026-08-29
- First seen: 2026-08-15
Recent updates
Self-evolving Cordis candidates; real-Loader admission; Harbor evaluation; auditable lineage; crash-resumable; trusted-base separation; fixed-replay effect verified.
FAQ
- Is a candidate trusted?
- No — every candidate is treated as untrusted and only accepted when source identity, evidence, cost, lifecycle, and recovery path reconcile against the trusted base.
- Does v0.2.0 claim benchmark wins?
- No — it verifies stable iteration and a measurable fixed-replay engineering effect; it does not claim Terminal-Bench/leaderboard/SOTA results.
- How do I install it?
- dsh plugin --profile headless add @dsh-self-evolving/core@0.2.3 (npm), or the source setup for development.