[jeffdupont]

Summary

Fixes AIC-3592 (1.0 blocker) in TypeScript. toLangGraph failed every turn that handed off: the Whodunit app's check --engine langgraph failed 30 of 30 graph turns with "Node whodunit-manor-butler is not reachable". Implements the new §2.x.3 bullets in launchdarkly/ai-sdks-monorepo#118.

In packages/langchain-agents/src/native-graph.ts:

• The hand-off Command answers its tool call. It now adds a ToolMessage (Handed off to <target>.) with the call's id, so the target's model call never sees an unanswered tool_use.
• The tools node declares its hand-off targets as ends, so LangGraph accepts the Command target. This is the "not reachable" error.
• The routing node does not run again after a hand-off. The static tools -> node edge is replaced by a conditional edge that goes back to the node only when no hand-off routed. This applies to single-child and multi-child nodes.
• Only the first hand-off in a message routes, as graph() does. A later one returns Ignored: only the first transfer in a turn routes. and does not track $ld:ai:graph:handoff_success.
• parallel_tool_calls is no longer passed to bindTools. Bedrock Converse rejects it; the first-hand-off rule replaces it.
• A hand-off target sees provider-neutral text history. Messages already in the state when a node is entered are rewritten for its model: reasoning and other non-text blocks dropped, each assistant message's tool calls appended as [Tool call] <name> <args> lines, the tool results after it folded into one user message of [Tool result] <name>: <text> lines in call order. Human and system messages pass unchanged. The node's own tool loop after entry stays native.

A normal tool call in the same message as a hand-off still runs, in the same tools step, and the target receives its result (decision 1A on the ticket).

Tests

native-graph.shallow.test.ts runs six new cases on the real StateGraph with a mock model: a multi-child hand-off, a single-child hand-off, a normal call plus a hand-off, two hand-offs, no parallel_tool_calls, and a node's own tool loop staying native. Each hand-off case asserts the exact messages the target receives (no tool role, tool calls, or reasoning blocks). A toNeutralHistory unit test covers every rewrite rule on the same input as the Python test, with the same expected output. The helper is exported from native-graph.ts for that test only; the package index does not re-export it (api:check passes). The unused _text and the duplicate final-output extraction now use contentText. The mocked executor test now asserts the ToolMessage in the Command update.

Local: yarn code:check, yarn typecheck, yarn api:check, yarn test all pass. The live Bedrock run is under "Live check".

Python: launchdarkly/python-ai-sdk#246. Go is out of scope for 1.0.

Live check

Passed on 29 Oct at head 0a0dd22: Whodunit check --engine langgraph on Bedrock completed 15 of 15 graph lines with 0 errors (the run at b528263 completed 9 of 16, with the toolConfig and reasoning_content errors). Reachability 5/5, verdict format 6/6. The later commit ae7a03e changes only the test file.

Bedrock end-to-end in CI: deferred to Q4 follow-up.

🤖 Generated with Claude Code

---

Note

Overview
Fixes broken LangGraph hand-offs in toLangGraph that caused unreachable target nodes and failed multi-agent runs (e.g. Whodunit on Bedrock).

Hand-off tools now return a Command that includes a ToolMessage answering the transfer call, register ToolNode ends for valid goto targets, and use conditional edges so the routing node does not run again after a successful hand-off. Only the first transfer_to_* in a turn routes; later ones return an ignored message. parallel_tool_calls is no longer passed to bindTools (Bedrock compatibility); first-wins routing replaces it.

When a node is entered via hand-off, prior conversation is passed through new toNeutralHistory (reasoning/non-text stripped, tool calls/results as plain [Tool call] / [Tool result] lines); the node’s own tool loop after entry stays native.

Adds real StateGraph integration tests for hand-off scenarios and a toNeutralHistory unit test aligned with Python; updates the mocked hand-off executor test for the Command message update.

^{Reviewed by Cursor Bugbot for commit ae7a03e. Bugbot is set up for automated code reviews on this repo. Configure here.}