Note

Medium Risk
Introduces new agent-graph APIs and changes the tracking model from a single tracker instance to per-invocation createTracker() factories with resumable tokens, which is a behavior/API change across the SDK. While covered by extensive new tests, it touches core metrics/judge/chat flows and could affect event correlation or downstream integrations.

Overview
Adds agent graph support to @launchdarkly/server-sdk-ai via LDAIClient.agentGraph(), introducing AgentGraphDefinition/AgentGraphNode with forward and reverse BFS traversal, validation (enabled/root/reachability/all agents enabled), and graph-level telemetry through a new LDGraphTracker (including resumption tokens).

Refactors AI config tracking to be per-invocation: configs now expose createTracker() instead of a long-lived tracker, LDAIConfigTracker gains runId, URL-safe resumptionToken, and at-most-once semantics for key metrics; adds createTracker(token, context) to reconstruct trackers and adds tool-call tracking (trackToolCall(s)).

Updates judge/chat flows to use the new tracker model and a simplified judge output shape (LDJudgeResult with score/reasoning/metricKey + sampled/error handling), updates examples accordingly, and adds comprehensive unit tests plus a new agent-graph-traversal example workspace.

^{Reviewed by Cursor Bugbot for commit ba595b1. Bugbot is set up for automated code reviews on this repo. Configure here.}