1052326311/dsh-plan-lattice

1052326311★ 1TypeScript最后同步: 2026-08-15

在 GitHub 打开

Evidence-gated recursive work graphs for long-horizon DeepSeek Harness agents

README 摘要

dsh-plan-lattice Evidence-gated work graph for long-horizon DeepSeek Harness agents. Every plan mutation earns its context. No parent completes before its evidence reconciles. dsh-plan-lattice is an independent community plugin for DeepSeek Harness. It implements the "Fractal Ledger" method: begin with a small project tree, recursively split only the active frontier into atomic leaves, and make each meaningful change prove that the agent has just read the current project contract. It is designed for the failure mode where a long-running agent starts with a good plan, discovers new constraints during implementation, and quietly keeps executing an obsolete version of that plan. What It Enforces The plugin makes ten rules executable rather than advisory: 1. Configurable execution intake. Long work can run with intakeMode: off , adaptive , or guided . off preserves the original autonomous workflow. adaptive lets the user choose guided clarification or model-led autonomy for each long task. guided asks only high-impact missing questions and requires approval of the exact contract before planning starts. 2. Boundary before decomposition. Intake separates confirmed facts, product decision…

在 GitHub 查看完整 README →
Agent/智能体agent-planningcontext-engineeringdeepseek-harnessdsh-pluginlong-horizon-agentsagent

分类