fix(eve): keep approvals pending across a same-step runtime action

.changeset/hitl-runtime-action-collision.md

@@ -0,0 +1,5 @@
+---
+"eve": patch
+---
+
+A model step that requests a tool approval (or question) and a subagent or remote-agent call in the same response no longer drops the approval. The harness now parks on both: the input request surfaces immediately, the delegation runs, and when its result arrives the turn re-parks on the still-pending approval instead of calling the model with a dangling tool call (`AI_MissingToolResultsError`).

packages/eve/src/harness/tool-loop.test.ts

@@ -913,6 +913,80 @@ describe("createToolLoopHarness", () => {
     ]);
   });
 
+  it("parks on both batches when one step carries a runtime action and an approval", async () => {
+    const gateToolCall = {
+      input: { action: "run" },
+      toolCallId: "gate-1",
+      toolName: "add",
+      type: "tool-call" as const,
+    };
+    const delegateToolCall = {
+      input: { message: "probe" },
+      toolCallId: "delegate-1",
+      toolName: "delegate",
+      type: "tool-call" as const,
+    };
+    setupMockAgent({
+      content: [
+        gateToolCall,
+        { approvalId: "approval-1", toolCallId: "gate-1", type: "tool-approval-request" },
+        delegateToolCall,
+      ],
+      finishReason: "tool-calls",
+      response: {
+        messages: [
+          {
+            content: [
+              gateToolCall,
+              { approvalId: "approval-1", toolCallId: "gate-1", type: "tool-approval-request" },
+              delegateToolCall,
+            ],
+            role: "assistant",
+          },
+        ],
+      },
+      text: "",
+      toolCalls: [gateToolCall, delegateToolCall],
+      toolResults: [],
+    });
+
+    const { emit, events } = createEventCollector();
+    const runStep = createToolLoopHarness(
+      createTestConfig("conversation", emit, { tools: createDelegationToolMap() }),
+    );
+
+    const parked = await runStep(createTestSession(), { message: "Gate and delegate." });
+
+    expect(parked.next).toBeNull();
+    expect(getPendingRuntimeActionBatch(parked.session.state)?.actions).toEqual([
+      expect.objectContaining({ callId: "delegate-1", kind: "subagent-call" }),
+    ]);
+    expect(hasPendingInputBatch(parked.session.state)).toBe(true);
+    expect(events.filter((event) => event.type === "input.requested")).toHaveLength(1);
+
+    // Runtime action results resume the step; the unanswered approval must
+    // re-park with the resolved results committed to history — not reach the
+    // model with a dangling `gate-1` tool call.
+    const reparked = await runStep(parked.session, {
+      runtimeActionResults: [
+        {
+          callId: "delegate-1",
+          kind: "subagent-result",
+          output: "delegated-done",
+          subagentName: "worker",
+        },
+      ],
+    });
+
+    expect(reparked.next).toBeNull();
+    expect(getPendingRuntimeActionBatch(reparked.session.state)).toBeUndefined();
+    expect(hasPendingInputBatch(reparked.session.state)).toBe(true);
+    const toolMessages = reparked.session.history.filter((message) => message.role === "tool");
+    expect(JSON.stringify(toolMessages)).toContain("delegated-done");
+    expect(events.filter((event) => event.type === "subagent.completed")).toHaveLength(1);
+    expect(events.at(-1)?.type).toBe("session.waiting");
+  });
+
   it("publishes declared subagent calls from deeply nested sessions", async () => {
     setupMockAgent({
       finishReason: "tool-calls",

packages/eve/src/harness/tool-loop.ts

@@ -585,6 +585,29 @@ export function createToolLoopHarness(config: ToolLoopHarnessConfig): StepFn {
         };
       }
 
+      // A step that parked on both runtime actions and input requests
+      // resumes here with the action results resolved into `messages` but
+      // the input batch still unanswered. Commit the resolved messages
+      // before re-parking — parking on the untouched session would drop
+      // the action results from history — and close the turn like the
+      // plain input park does, so clients see the waiting boundary.
+      if (resolvedRuntimeActions.outcome === "resolved") {
+        let parkedSession: HarnessSession = {
+          ...pending.session,
+          history: [...resolvedRuntimeActions.messages],
+        };
+        if (emit && config.mode === "conversation") {
+          emissionState = await emitTurnEpilogue(
+            emit,
+            emissionState,
+            config.mode,
+            parkedSession.continuationToken,
+          );
+          parkedSession = setHarnessEmissionState(parkedSession, emissionState);
+        }
+        return { next: null, session: parkedSession };
+      }
+
       return { next: null, session: pending.session };
     }
 
@@ -1905,21 +1928,50 @@ async function handleStepResult(input: {
     // parked session carries the default emission state (turnId ""),
     // because the post-preamble `setHarnessEmissionState` is dropped by
     // the later `session = pending.session` / `maybeCompact` rebinds.
-    return {
-      next: null,
-      session: setHarnessEmissionState(
-        setPendingRuntimeActionBatch({
-          actions: pendingRuntimeActions,
-          event: {
+    let parkedSession = setPendingRuntimeActionBatch({
+      actions: pendingRuntimeActions,
+      event: {
+        sequence: emissionState.sequence,
+        stepIndex: emissionState.stepIndex,
+        turnId: emissionState.turnId,
+      },
+      responseMessages,
+      session: { ...baseSession, history: [...promptMessages] },
+    });
+
+    // The same step may also carry approval or question requests. Persist
+    // them alongside the runtime-action batch — with empty
+    // `responseMessages`, since the action batch owns the step's assistant
+    // messages — and surface them now so channels collect the answer while
+    // the actions run. The resumed step resolves the action results first,
+    // then re-parks on this batch (see `executeStepBody`).
+    if (inputRequests.length > 0) {
+      parkedSession = setPendingInputBatch({
+        event: {
+          sequence: emissionState.sequence,
+          stepIndex: emissionState.stepIndex,
+          turnId: emissionState.turnId,
+        },
+        requests: inputRequests,
+        responseMessages: [],
+        session: parkedSession,
+      });
+
+      if (emit) {
+        await emit(
+          createInputRequestedEvent({
+            requests: inputRequests,
             sequence: emissionState.sequence,
             stepIndex: emissionState.stepIndex,
             turnId: emissionState.turnId,
-          },
-          responseMessages,
-          session: { ...baseSession, history: [...promptMessages] },
-        }),
-        emissionState,
-      ),
+          }),
+        );
+      }
+    }
+
+    return {
+      next: null,
+      session: setHarnessEmissionState(parkedSession, emissionState),
     };
   }