[XieX]

What

AIC-3599. In graph().stream(), a node's own judges ran and their scores went to OpenTelemetry, but never reached the stream. stream_node and stream_route both put judge_results into the router's outcome, yet the node_done event they yield had only type, nodeKey, response and usage. The done event's judgeResults holds only the graph-level judge, so a caller reading scores from the stream got nothing for a per-node judge.

Change

• node_done now carries judgeResults for that node's own judges. It is omitted when none of them produced a result, the same rule done uses, and it is built the same way done_event is.
• done is unchanged and still holds only the graph-level judge (graph_judge). Merging per-node results into it would put them in one dict keyed by judge key, so two nodes using the same judge would overwrite each other. A caller that wants every score reads each node_done.
• The GraphStreamEvent shape comment in types.py lists the new field, and examples/graph_streaming.py prints a node's judgeResults on its [node_done] line when present.

No public signature changes, so the API reports are unchanged.

Tests

• test_node_done_carries_the_nodes_own_judge_results: a two-node graph where only the leaf's judge returns a score. The leaf's node_done carries it, the root's has no judgeResults key, and done has none.
• test_node_judges_stay_on_node_done_when_a_graph_judge_runs: the same graph with graph_judge set as well. The leaf's node_done keeps its own result and done holds only the graph judge's.
• test_routing_node_judge_results_on_its_node_done: the same for a node that routes through a handoff tool.

All three fail without the fix. make lint, make format-check, make typecheck, make api-report-check and make test pass.

Not in this PR

graph().invoke() still returns only the graph-level judge's results. Exposing per-node results there needs a new response shape and is tracked separately.

JavaScript and Go get the same fix in their own PRs.

🤖 Generated with Claude Code

---

Note

Overview
Fixes a gap where per-node judges ran during graph().stream() but their scores never appeared on the event stream—only graph-level results on the final done event were visible.

node_done now optionally includes judgeResults for that node’s own judges (same omission rule as done when nothing scored). This is applied in both normal node completion and routing/handoff paths in graph.py. The final done event is unchanged and still carries only the graph judge, so callers that need every score should read each node_done.

Documentation and the graph streaming example are updated to show the new field; tests cover leaf nodes, graph judges running alongside node judges, and routing nodes.

^{Reviewed by Cursor Bugbot for commit 579b255. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 12328fb. Configure here.}

[jeffdupont]

LGTM. Checked node_done judgeResults against the spec in ai-sdks-monorepo#111 and the merged JS change (js-ai-sdk#182): omitted when empty, done stays graph-judge only. The 3 new tests fail with graph.py reverted to main and all 42 stream tests pass on the branch.