[jeffdupont]

Fixes AIC-3571.

Why

reverseTraverse is meant to visit every child before its parent, so adapters can build children first. It walked up breadth-first from the terminal nodes instead, so a parent was visited as soon as one child was. With root → P, P → T1, P → Q, Q → T2 the order was T1, T2, P, Q, root: P came before its child Q. toOpenAIAgents then threw Child agent "Q" was not built before parent "P", and toClaudeAgents built P without Q's tool and said nothing. The Python adapters already use a post-order DFS (_reverse_traverse in claude-agents, _visit in openai-agents); this PR makes core match.

Fix

• packages/client/src/graph.ts: reverseTraverse is now a post-order DFS from the root with a visited set. Each node is visited once, every child before its parent, root last. Same signature.
• packages/openai-agents/src/native-graph.ts: a child that is still unbuilt (only possible on a cycle back-edge now) is dropped with a console.warn instead of throwing.
• packages/claude-agents/src/native-graph.ts: the same case used to be dropped silently by .filter(Boolean). It now logs a console.warn.
• Doc comments in types.ts and AGENTS.md no longer say "reverse BFS".

Behaviour changes:

• Nodes that can't be reached from the root are no longer visited.
• A graph with no terminal nodes (a pure cycle) is now walked from the root. It used to return undefined without visiting anything.

Testing

• yarn build, yarn typecheck, yarn code:check, yarn api:check (client): all pass.
• yarn test: 2335 passed, 11 skipped, 0 failed across 11 suites (openai-agents: 128 passed).
• New tests: in client, the uneven-depth order (expects T1, T2, Q, P, root) and a cycle visited once each. In openai-agents, the uneven graph built through the real resolveGraph and reverseTraverse (no mock), plus the cycle back-edge dropped with a warning. In claude-agents, the cycle back-edge warning.
• With the fix reverted, these fail: the client uneven test (Q out of order), the openai-agents uneven test (Child agent "Q" was not built before parent "P"), and the claude-agents cycle test (no warning). The client cycle test also passes on the old code. It stays as a guard that each node is visited once.
• The openai-agents OTel mock gains setGlobalTracerProvider: vi.fn() because the real initClient (BYOC path) registers a tracer provider.

Related: js #147 · python launchdarkly/python-ai-sdk#189 · go launchdarkly/go-ai-sdk#12 · spec https://github.com/launchdarkly/ai-sdks-monorepo/pull/73

🤖 Generated with Claude Code

---

Note

Overview
reverseTraverse in the core client no longer walks upward breadth-first from leaves. It now does a post-order DFS from the root: each node is visited once, every descendant before its parent, root last. That matches Python adapters and fixes uneven graphs where a parent was visited before a deeper child (e.g. P before Q), which broke toOpenAIAgents and left toClaudeAgents with missing child tools.

Semantics shift: only nodes reachable from the root are visited; a graph with no leaves but a cycle from the root is traversed instead of returning without visiting anything.

Native graph adapters (toOpenAIAgents, toClaudeAgents) drop cycle back-edges when the child agent/tool was not built yet, emitting [LaunchDarkly] Dropping cyclic edge "…" -> "…" via console.warn instead of throwing (OpenAI) or failing silently (Claude).

Docs/comments in types.ts and AGENTS.md describe post-order rather than “reverse BFS.” Tests cover uneven depth order, single visit in cycles, and end-to-end adapter builds with real resolveGraph.

^{Reviewed by Cursor Bugbot for commit 2fb274a. Bugbot is set up for automated code reviews on this repo. Configure here.}