dsh-web
zhu1090093659/dsh-webDeepSeek Harness(DSH)Web 插件聚合生态包 · 一切皆插件,创意工坊分发(dsh-market.com)
The unified kernel (kernel8 · ayeOS · cordis · mem8) lives under enthea v108; THIS harness is its N+1: the sovereign, agent-driven evaluation lane that watches the lane. Cordis plugin architecture + EntheAI backends + multi-model benchmarking — the examiner side of the same kernel. The corpus above all; the evaluations row the verdicts.
English | 中文
Deepsiper Enthea (deepsiper-enthea) is a sovereign, agent-driven LLM evaluation harness forked from deepseek-ai/deepseek-harness (dsh 0.1.0-rc.7). It provides end-to-end multi-model orchestration, self-hosted and sovereign backend integration, Cordis-powered extensible plugin pipelines, and JSON-RPC automation for modern LLM evaluation workflows.
goToDefinition, findReferences, goToImplementation, hover) shipped in the dsh-base core bundle with a TypeScript/JavaScript stdio provider, so the lsp tool is available in every profile.tool-eval, eval-entheai, and custom benchmarking metrics.Everything in Deepsiper Enthea is a composable Cordis plugin.
┌─────────────────────────────────────┐
│ OpenCode / JSON-RPC / CLI / Web │
└──────────────────┬──────────────────┘
│
┌──────────────────▼──────────────────┐
│ Cordis Kernel (Context & DI) │
└────┬──────────────┬───────────────┬─┘
│ │ │
┌──────────────▼─────┐ ┌──────▼──────┐ ┌──────▼──────────────┐
│ Sovereign Backends│ │ Eval Plugins│ │ Sandboxed Tool Seams│
│ (EntheAI / Local) │ │ (tool-eval) │ │ (Landlock / Bash) │
└────────────────────┘ └─────────────┘ └─────────────────────┘
>=22.19.0 or >=24.0.0, TypeScript 6 (Strict ESM)tsdown / rolldown + tsc project references^22.19.0 || >=24.0.0pnpm >= 11.0.0# Clone repository
git clone https://github.com/8b-is/deepsiper-enthea.git
cd deepsiper-enthea
# Install dependencies and build harness
pnpm install
pnpm build
Deepsiper Enthea includes an automated multi-case AST evaluation harness (examples/eval-entheai/driver/benchmark_sweep.py) verifying mathematical precision, algorithm syntax, and zero reward-hacking across isolated execution environments:
| Benchmark Task | Category | Trials | Pass Rate | Mean Latency |
|---|---|---|---|---|
| FizzBuzz Logic & 18 Edge Cases | Logic Verification | 2 / 2 | 100.0% | ~16.8 s |
| $O(\log N)$ Matrix Power Fibonacci | Mathematical Exponentiation | 2 / 2 | 100.0% | ~16.1 s |
| Kademlia 256-bit XOR Metric Distance | Distributed DHT Routing | 2 / 2 | 100.0% | ~12.8 s |
| BitLinear ${-1, 0, +1}$ Quantizer | Ternary Weight Mapping | 2 / 2 | 100.0% | ~13.9 s |
| AST Invariant & Pure Syntax Validator | Syntax Tree Verification | 2 / 2 | 100.0% | ~13.6 s |
# Keyless fast gate (no dsh runtime): grading logic + canonical validator
pnpm bench:entheai:verify
# Full multi-case sweep against the built dsh runtime (inline stub backend)
pnpm bench:entheai
https://axiomquant.orghttps://github.com/8b-is/classroom-sota-traininghttps://github.com/peterlodri-sec/etherhive · https://etherhive.vaked.devhttps://vaked.devhttps://peterl.dev@0xp3t3rl.bsky.socialMIT © 8b-is & DeepSeek AI contributors. Third-party dependency notices are listed in THIRD_PARTY_NOTICES.md.
Genesis Seal: 7c242080f5f821e5eaf563fe2208d60632c451687baf65f4fe8e4a0d226e3ecf · WE. {-1, 0, +1}. <3