fix(react-headless): order LangGraph multi-step tool loops correctly

[AbhinRustagi]

What

Two bugs in the same tool-loop path — one in how LangGraph streams are decoded, one in how any provider-executed tool activity is rendered.

Stream. The LangGraph adapter assumed one assistant message per stream. A graph that called a tool and then resumed the model collapsed into a single merged message with no tool results, so what you saw live didn't match what the next request replayed.

Render. A turn whose tool activity is all provider-executed crashed the thread with can't access property "type", current is undefined. An OpenUI Cloud scaffold reproduces it.

Changes

react-headless — LangGraph adapter

• Track the current message id per graph step, so each model step becomes its own assistant message.
• Handle ToolMessage / ToolMessageChunk — a result closes the open tool call and the assistant segment before emitting TOOL_CALL_RESULT, keeping history ordered ai → tool → ai.
• Prefer tool_call_chunks over tool_calls when both are present. LangChain chunks carry a provisional tool_calls projection, often with empty args, which duplicated calls.
• Track open and started tool calls separately so each call ends exactly once.
• Mark status: "error" tool messages as errored results.

react-ui — timeline rendering

• Thread built its rows only from tool calls it could match back to an assistant segment, silently discarding the orphan activities pairToolActivity synthesizes for tool results whose call lives on no assistant message. Provider-executed tools are exactly that shape, so their activity never rendered — and when a turn had only orphans, steps came back empty while turnActivities did not.
• ToolCallTimeline guarded on activities but indexed steps, so that combination evaluated steps[-1] and threw. Unclaimed activities are now appended as rows, and the guard covers the array actually indexed.

Test Plan

• Verified locally

react-headless 168/168 and react-ui 29/29 passing; prettier and tsc clean. Adds unit coverage plus an integration test asserting a tool-loop turn stays ordered and valid for the next request.

The render fix is verified by the existing suites plus tracing the failure end to end — pairToolActivity synthesizes orphans (already covered by its own tests), Thread dropped them, ToolCallTimeline then indexed [-1]. It has no direct regression test, since this repo's convention is not to add test files unprompted.

Checklist

• I considered backwards compatibility

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
openui-docs	Ready	Preview	Aug 14, 2026 11:20am