[XieX]

Summary

The text for graph nodes with more than one outgoing edge told every such node to "Finish your own work and provide your final response". It also said to call a transfer tool "Only once you have your final answer". A pure router has no answer of its own, so this wording invites it to write the target agent's answer. In stream() that reply reaches the caller as chunk events, and it reaches the target as the [Previous agent response] block. In the Whodunit test app the router answered in another agent's voice. The wording is a likely cause, not a proven one.

Both strings in build_handoff_routing now match launchdarkly/ai-sdks-monorepo#104 (TESTING.md §3.15a) byte for byte. TypeScript and Go get the same change in separate PRs.

• Routing suffix (appended to the node's instructions): \n\nComplete your task using your available tools first. Only once your own part is done, call exactly one transfer tool to route to the next agent. Do not answer the parts of the request that belong to the agent you route to.
• Handoff tool result: Handoff to <target> recorded. <target> takes over when your turn ends. Finish your own part, then end your turn. Do not answer for <target>. (<target> is the raw target_key, not the sanitized tool name.)

The new text keeps recorded and using your available tools first. Those words are why a node with a real tool still calls it before routing (#90). The comment above the handoff function now gives both constraints. The text does not branch on whether a node is a router, because the graph schema has no router role. No other routing behaviour changes.

Tests

TestGraphStreamMultiEdge::test_routing_text_matches_the_spec checks both strings exactly through graph().invoke() and graph().stream(). It covers a multi-edge root that has its own tool and one that has none. The assertions read the config and tool handlers the handler actually received: the instructions, the result each __handoff_* handler returns, and the set of tools. With the old strings, all four cases fail.

Client suite: 1907 passed, 1 skipped. ruff check, ruff format --check, mypy packages/*/src and the API report check are all clean.

packages/langchain-agents/tests/test_handler.py::test_handoff_handler_records_destination keeps its own copy of the old result text. It is a local fixture that only checks passthrough, so I left it alone here. Changing it would also bump that package.

Live checks not run

The spec asks for two live checks on one model before a release ships this. Neither has been run:

1. In a two-edge graph whose root has no tools, the root does not answer for the agent it picks.
2. In a two-edge graph whose root has a real tool, the root still calls that tool before routing.

🤖 Generated with Claude Code

---

Note

Overview
Updates multi-edge graph routing prompts in build_handoff_routing so routers are not told to deliver a “final answer/response,” which was encouraging pure routing nodes to speak for the next agent (visible in stream chunks and in the [Previous agent response] handoff block).

The instruction suffix now tells the model to finish its own part, call one transfer tool, and not answer portions of the request owned by the routed agent. The __handoff_* tool return string is rewritten to say the target takes over when the turn ends and explicitly not to answer for that target. Both strings match TESTING.md §3.15a byte-for-byte across SDKs; the inline comment documents why recorded and the no–final-response wording stay.

Adds test_routing_text_matches_the_spec, asserting exact copy through graph().invoke() and graph().stream() for roots with and without their own tools.

^{Reviewed by Cursor Bugbot for commit deebf59. Bugbot is set up for automated code reviews on this repo. Configure here.}

[jeffdupont]

Approving. Both strings match the spec in launchdarkly/ai-sdks-monorepo#104 exactly, and the new test fails on the old wording. Tests, typecheck, lint, format and the API report check all pass. Please merge after the spec PR.

Small follow-ups, none blocking:

• packages/langchain-agents/tests/test_handler.py:523-533 still has the old wording in a fixture. Worth updating in a follow-up.
• The comment at graph.py:340-347 could be a short pointer to the spec.