fix(cli): report usage for steps that end after run goes idle (#60475)

packages/cli/src/run/noninteractive.ts

@@ -85,6 +85,9 @@ export async function runNonInteractivePrompt(input: Input) {
   const renderedText = new Map<string, string>()
   const renderedReasoning = new Map<string, string>()
   const renderedTools = new Set<string>()
+  // Live step events stop being read once the session goes idle, and the last step usually ends after
+  // that. Reconcile uses this to emit only the step_start/step_finish lines the live stream missed.
+  const steps = new Map<string, "started" | "ended">()
   let submitted = false
   let promoted = false
   let emittedError = false
@@ -280,6 +283,7 @@ export async function runNonInteractivePrompt(input: Input) {
           }
           continue
         }
+        steps.set(event.data.assistantMessageID, "started")
         if (!emit("step_start", time, { part }) && input.format !== "json") {
           UI.empty()
           UI.println(`> ${event.data.agent} · ${event.data.model.id}`)
@@ -493,6 +497,7 @@ export async function runNonInteractivePrompt(input: Input) {
 
       if (event.type === "session.step.ended") {
         flushStep()
+        steps.set(event.data.assistantMessageID, "ended")
         const part = {
           id: partID(event.id),
           sessionID: input.sessionID,
@@ -568,6 +573,18 @@ export async function runNonInteractivePrompt(input: Input) {
     for (const message of projected.messages) {
       if (message.type !== "assistant") continue
       const timestamp = message.time.completed ?? message.time.created
+      const step = steps.get(message.id)
+      if (!step) {
+        emit("step_start", message.time.created, {
+          part: {
+            id: projectedPartID(message.id, "step-start"),
+            sessionID: input.sessionID,
+            messageID: message.id,
+            type: "step-start",
+            snapshot: message.snapshot?.start,
+          },
+        })
+      }
       let textOrdinal = 0
       let reasoningOrdinal = 0
       for (const item of message.content) {
@@ -660,6 +677,21 @@ export async function runNonInteractivePrompt(input: Input) {
         UI.error(item.state.error.message)
       }
 
+      if (step !== "ended" && message.finish && message.tokens) {
+        emit("step_finish", timestamp, {
+          part: {
+            id: projectedPartID(message.id, "step-finish"),
+            sessionID: input.sessionID,
+            messageID: message.id,
+            type: "step-finish",
+            reason: message.finish,
+            snapshot: message.snapshot?.end,
+            cost: message.cost ?? 0,
+            tokens: message.tokens,
+          },
+        })
+      }
+
       // A declined tool call ends its step with an interrupted-step error that is
       // only a consequence of our own rejection; it was already reported above.
       if (message.error && !emittedError && !permissionRejected && !formCancelled) {

packages/cli/test/run/noninteractive.test.ts

@@ -373,7 +373,51 @@ describe("runNonInteractivePrompt", () => {
         .split("\n")
         .filter(Boolean)
         .map((line) => JSON.parse(line)),
-    ).toEqual([expect.objectContaining({ type: "text", part: expect.objectContaining({ text: "projected answer" }) })])
+    ).toEqual([
+      expect.objectContaining({ type: "step_start", part: expect.objectContaining({ messageID: "msg_assistant" }) }),
+      expect.objectContaining({ type: "text", part: expect.objectContaining({ text: "projected answer" }) }),
+    ])
+  })
+
+  test("reports usage for a step that ends after the Session goes idle", async () => {
+    const tokens = { input: 3, output: 5, reasoning: 7, cache: { read: 11, write: 13 } }
+    const assistant = (messageID: string): SessionMessageInfo[] => [
+      {
+        id: "msg_assistant",
+        type: "assistant",
+        agent: "build",
+        model: { providerID: "test", id: "test-model" },
+        content: [{ type: "text", text: "done" }],
+        finish: "stop",
+        cost: 0.25,
+        tokens,
+        time: { created: 2, completed: 3 },
+      },
+      { id: messageID, type: "user", text: "hello", time: { created: 1 } },
+    ]
+    const lines = async (turn: (messageID: string) => V2Event[]) =>
+      (await capture({ format: "json", turn, messages: assistant, wait: () => Promise.resolve() })).stdout
+        .split("\n")
+        .filter(Boolean)
+        .map((line) => JSON.parse(line))
+
+    const missed = await lines((messageID) => [prompted(messageID), stepStarted()])
+    expect(missed.map((event) => event.type)).toEqual(["step_start", "text", "step_finish"])
+    expect(missed[2].part).toMatchObject({ messageID: "msg_assistant", reason: "stop", cost: 0.25, tokens })
+
+    const seen = await lines((messageID) => [
+      prompted(messageID),
+      stepStarted(),
+      {
+        id: "evt_step_ended",
+        created: 2,
+        type: "session.step.ended",
+        durable: { aggregateID: "ses_1", seq: 2, version: 1 },
+        data: { sessionID: "ses_1", assistantMessageID: "msg_assistant", finish: "stop", cost: 0.25, tokens },
+      },
+      settled(),
+    ])
+    expect(seen.filter((event) => event.type === "step_finish")).toHaveLength(1)
   })
 
   test("reports an observed execution failure before prompt promotion", async () => {