dsh-graph-monitor

Debugging a multi-step AI pipeline often means staring at logs and guessing which branch the model actually took. dsh-graph-monitor brings a LangGraph-style graph workflow into DeepSeek Harness: you describe nodes, edges, conditional routing, parallel branches, and structured State, and as execution runs the SVG topology animates in real time—the current node pulses purple, finished ones turn green, failures glow red, and the traversed edge highlights in amber. The graph is the engine's actual execution topology, so what you see is exactly what runs.

It is far more than a static DAG viewer. Six control-flow primitives—gate, switch, subgraph, loop, retry, timeout—let you nest sub-graphs, iterate, retry on error, and enforce time budgets. LLM and Agent nodes can also emit structured graph-editing ops so the model itself drives orchestration, with atomic validation before application. The execution backend offers two modes: simulated for pure animation demos, and real for binding to DSH's actual fn, tool, agent, and llm capabilities with automatic fallback when a capability is missing. Trigger workflows in-chat via /workflow or let the model invoke graph_monitor_run with natural language, and revisit any past run in the history panel.

Built for DSH developers assembling multi-step agent pipelines who need visual orchestration and live debugging, and for teams that want natural-language-driven workflows with a clear, real-time view of every branch and its outcome.