OpenCnid/deepseek-rlm

OpenCnid★ 2TypeScript最后同步: 2026-08-22

在 GitHub 打开

Persistent IPython state, durable snapshots, and native recursive subagents for DeepSeek Harness—packaged as a Prime-compatible Cordis plugin bundle.

README 摘要

DeepSeek RLM Give every DeepSeek Harness agent a persistent IPython workspace and the ability to delegate focused work to native recursive subagents. DeepSeek Harness remains the only agent loop: it owns models, tools, approvals, sessions, policy, lineage, cancellation, and persistence. [!NOTE] RLM means a Recursive Language Model workflow. This is a Cordis plugin bundle for DeepSeek Harness—not another harness and not a DeepSeek model. [!IMPORTANT] This preview targets the exact DSH 0.1.0-rc.7 baseline and requires three ordered host patches. The packages are not published; install the local tarballs using the installation guide. What it adds - Persistent Python: one authenticated IPython/Jupyter kernel per live DSH agent, with state preserved between tool calls. - Durable snapshots: bounded, digest-authorized namespace snapshots restore variables without replaying historical cells or their side effects. - Native recursion: await rlm(...) admits a normal DSH continuable child with enforced lineage, depth, model selection, and cancellation. - Parent/child messaging: Prime-compatible agent message calls route through public DSH subagent services. - Harness tools from Python: optiona…

在 GitHub 查看完整 README →
Agent/智能体agent-orchestrationagentic-aiai-agentscordiscordis-plugindeepseek-harnessdshdsh-plugin

分类