issuegraph

A GitHub issue triage agent built on LangGraph. It classifies an issue, routes it to a specialist drafter, loops the reply through a quality guard, and pauses for human approval when its own confidence is low. This page runs the real graph, streaming each node as it executes.

pick an issue
the graph, live
classify
category + confidence
draft
specialist reply node
guard
LLM judge: good enough?
gate
confident? finalize : ask a human

Pick a preset above. You will watch the LangGraph state machine run node by node: classify, route to a specialist drafter, loop through a quality guard, then either finalize or stop and ask you to approve. The last preset is intentionally ambiguous and usually pauses the graph for human review.

evals + calibration
100%
category accuracy
100%
draft quality (LLM judge)
0.003
brier score
8
golden examples
confidence bucketnclaimedactualverdict
0.9-1.080.951.00calibrated

Numbers from the committed golden-set run (8 labeled issues) scored two ways: deterministic exact-match on the category, and an LLM judge on the drafted reply. The Brier score and reliability table check whether the classifier's stated confidence is honest. Small clean golden sets flatter the model; the interesting work is growing the set with production failures.