dsh-graph-monitor

Debugging a multi-step AI pipeline usually means scrolling through logs and guessing which branch the model actually followed. dsh-graph-monitor brings a LangGraph-style graph workflow into DeepSeek Harness: define nodes, edges, conditional routing, parallel branches, and shared State, then watch the SVG topology light up as the engine executes—purple pulse on the active node, green on completion, red on failure, amber highlight on the traversed edge. The diagram is the engine's real execution graph, so there is zero drift between what you see and what runs.

Beyond a static DAG viewer, it ships six control-flow primitives—gate, switch, subgraph, loop, retry, and timeout—that let you nest sub-graphs, iterate until a predicate fails, retry on error, and enforce a time budget. LLM and Agent nodes can emit structured graph-mutation ops, so the model itself decides orchestration, validated atomically before application. The backend supports simulated mode for pure animation demos and real mode that binds fn, tool, agent, and llm nodes to DSH's actual capabilities, falling back gracefully when a capability is missing. Kick off workflows with /workflow in chat or let the model call graph_monitor_run with plain natural language, then replay any past run from the history panel.

If you are a DSH developer building multi-step agent pipelines and need visual orchestration with live debugging, or a team that wants natural-language-triggered workflows with a clear, real-time view of every branch and its outcome, dsh-graph-monitor is built for you.