fix(agent): close open steps before the step-limit reply finishes the run

[mxmzb]

Problem

When a Slack turn hits the graph step limit, the person sees

ChannelCanonicalRunError: Cannot send 'RUN_FINISHED' while steps are still active: TodoListMiddleware.after_model

instead of the "I hit the step limit… ask me again" reply.

The adapter opens a step for every graph node. When the graph raises GraphRecursionError it drops its run state without closing that step, so iter_agent_events in agent/agui.py sent RUN_FINISHED while the step was still open. @ag-ui/client's event verifier rejects that. A raw SSE read of the agent looks fine, because nothing on that side checks the event order.

Fix

iter_agent_events now tracks the steps, text messages, and tool calls the stream opens. On a step-limit crash it closes whatever is still open, innermost first, before the snapshots, the reply, and RUN_FINISHED.

Verification

• Reproduced by running a real adapter recursion run through verifyEvents from @ag-ui/client@0.0.59. Before the fix: Cannot send 'RUN_FINISHED' while steps are still active: step. After: accepted (30 events).
• Two new tests in agent/tests/test_agui_recursion.py, both failing before the fix:
  • the real adapter's step-limit path leaves nothing open at RUN_FINISHED;
  • a stream cut off with a step, a text message, and a tool call open gets them closed in order.
• uv run pytest (536 passed, 1 skipped), pnpm check-types, pnpm test (436 passed), Railway IaC (no diagnostics).
• Live in Slack (support, together with feat(arcade): Arcade as a second connected-app provider #73): the step-limit turn now ends with the polite reply instead of the error.

Not in this PR

The step budget itself is unchanged. A limit of 25 graph steps is about five tool calls, because each tool call costs four nodes (model, two after_model middlewares, tools). That's a separate decision.

🤖 Generated with Claude Code