Render a terminal result the streaming path did not recognise

[DavidNic11]

What & why

Every streaming turn on the Temporal engine fails in the UI with ❌ agent stream ended unexpectedly — including turns that fully succeeded.

Observed on a live cluster running AGENT_ENGINE=temporal: asked "who won the
world cup?", and the backend did everything right —

POST /invoke                                    202
ConversationWorkflow  CheckNeedsCapability → RetrieveSkills → RetrieveAgents
                    → SelectDelegate → ResolveSkillTools → PlanAction
                    → LaunchToolRun
ToolRun run-ad12da07…  Succeeded    Job Complete 1/1 in 5s

…and the user saw the error. Open WebUI streams, so this is every chat message.

The interface guarantees a result; the stream loop asked for a node name

AgentGraphLike is the seam ADR 0036 is built on, and TemporalEngine is a
second implementation of it. But every terminal branch in the streaming loop
keys off a LangGraph node name — composeResponse, planAction,
selectDelegate, bareAnswer, delegateToAgent, checkActiveAgentRun.

The engine has no LangGraph nodes. It yields one terminal update under its own
key, matches nothing, falls out of the loop, and hits the "shouldn't normally
happen" error.

Those branches are therefore unreachable for any implementation that is not the
graph — precisely the case the interface exists to allow.

Two checks in the same loop already got this right by reading update instead
of the node name: update.error and update.pendingToolCalls. So engine
failures and caller-supplied tool calls (ADR 0035) both worked. The success
path was the one left out.

The fix

Key the fallback on what the interface actually guarantees — if the stream ended
and a result was captured, render it.

Unreachable for the graph itself, which always hits a branch above first, so
AGENT_ENGINE=langgraph behaviour is unchanged.

Why not just name the engine's update after a graph node

Yielding { composeResponse: … } from TemporalEngine.stream() is a smaller
diff and would work. It would also mean a second implementation lying about
which node produced an answer, to satisfy a check that should not have been
node-specific — and it leaves the next implementation to rediscover this. Fixing
the consumer keeps the interface honest.

Verification

• apps/agent-orchestrator suite: 59 passed.
• The new engine-shaped test fails without the src change (confirmed by
  reverting it: 1 failed | 58 passed), so it genuinely covers the bug.
• Second test pins the existing behaviour: a stream ending with neither a
  terminal node nor a result still reports the error.

Deploying it

This needs an agent-orchestrator image rebuild, not just a chart bump — the
change is in app code. No chart or values change, so no version bump here.

🤖 Generated with Claude Code

[imaustink]

🚢