[XieX]

Summary

AIC-3601: when a graph node has more than one outgoing edge, the routing text told it 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 node that only routes 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. It also 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 of that, not a proven one.

This PR changes both strings in buildHandoffRouting (packages/client/src/graph.ts) to the wording pinned in the monorepo's TESTING.md §3.15a (launchdarkly/ai-sdks-monorepo#104, as of its latest commit c89e5f8). <target> is the edge's raw targetKey, not the sanitized tool name.

	Text
Routing suffix	\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>.

The text still says recorded and using your available tools first, so the fix from #59 holds: a node with a real tool calls it before routing. The text does not branch on whether a node is a router. The graph schema has no router role, and a node with no tools can still write an answer and then route. The comment above the handoff handler now points to §3.15a for the reasoning. Routing behaviour is otherwise unchanged. Python and Go are getting the same strings byte for byte in separate PRs.

Tests

• A new graph() multi-edge routing text block runs through both graph().invoke() and graph().stream(), for a root that has its own tool and for a root with no tools. It asserts exact strings for the instructions the handler received and for the value its __handoff_agent_b handler returned. It also asserts the tool and handler sets the handler received, and that the node's own tool still reaches the caller's handler.
• The existing stream parity test now expects the new handoff result.
• When graph.ts is reverted to main, these 5 tests fail.
• packages/client: yarn test passes (1929 passed, 11 skipped), yarn typecheck passes, and biome check is clean.

🤖 Generated with Claude Code

---

Note

Overview
Fixes AIC-3601 by replacing multi-edge routing copy in buildHandoffRouting so router nodes stop being told to deliver a “final answer” (which pushed them to speak for the next agent).

The instruction suffix now says to finish your own part before calling one transfer tool, and explicitly not to answer portions of the request meant for the routed agent. The __handoff_* tool return string now states the target takes over when the turn ends, to finish the router’s part and end the turn, and not to answer for the target. Wording matches the monorepo pin in TESTING.md §3.15a; <target> remains the raw edge targetKey.

Tests add a graph() multi-edge routing text suite (pinned constants) covering invoke() and stream() for roots with and without their own tools, and update an existing parity expectation for the new handoff message.

^{Reviewed by Cursor Bugbot for commit 3f03005. 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 tests fail on the old wording. Client tests and typecheck pass.

Small follow-ups, none blocking:

• The description cites spec commit 8284a10 and says live checks are required. The latest spec commit (1728f86) removed that, so please update it.
• The comment at graph.ts:261-266 could be a one-line pointer to the spec.