fix(provider-acp): Grok runs finish with their reply while background subagents keep working

apps/server/scripts/acp-mock-agent.ts

@@ -26,6 +26,13 @@ const emitInterleavedAssistantToolCalls =
 const emitV2Fidelity = process.env.T3_ACP_EMIT_V2_FIDELITY === "1";
 const vibeRetryOutcome = process.env.T3_ACP_VIBE_RETRY_OUTCOME;
 const emitGenericToolPlaceholders = process.env.T3_ACP_EMIT_GENERIC_TOOL_PLACEHOLDERS === "1";
+// Keeps the first prompt open after its generic placeholders until session/cancel,
+// so a test can interrupt a root that is still prompting.
+const holdFirstGenericPromptUntilCancel =
+  process.env.T3_ACP_HOLD_FIRST_GENERIC_PROMPT_UNTIL_CANCEL === "1";
+// Leaves a background monitor running when the generic prompt returns, so the
+// root settles held by background work.
+const emitRunningMonitorWithGeneric = process.env.T3_ACP_EMIT_RUNNING_MONITOR_WITH_GENERIC === "1";
 const emitPostSettleMonitorFlow = process.env.T3_ACP_EMIT_POST_SETTLE_MONITOR_FLOW === "1";
 const emitInTurnTaskOutputThenLateDuplicate =
   process.env.T3_ACP_EMIT_IN_TURN_TASKOUTPUT_THEN_LATE_DUPLICATE === "1";
@@ -1968,6 +1975,26 @@ const program = Effect.gen(function* () {
           },
         });
 
+        if (emitRunningMonitorWithGeneric && promptCount === 1) {
+          yield* agent.client.sessionUpdate({
+            sessionId: requestedSessionId,
+            update: {
+              sessionUpdate: "tool_call_update",
+              toolCallId: "tool-call-monitor-hold",
+              title: "Monitor: held background work",
+              kind: "execute",
+              status: "in_progress",
+              rawInput: {},
+            },
+          });
+        }
+        if (holdFirstGenericPromptUntilCancel && promptCount === 1) {
+          while (!cancelledSessions.has(requestedSessionId)) {
+            yield* Effect.sleep("25 millis");
+          }
+          cancelledSessions.delete(requestedSessionId);
+          return yield* finishPrompt(requestedSessionId, "cancelled");
+        }
         return yield* finishPrompt(requestedSessionId, "end_turn");
       }
 

apps/server/src/orchestration-v2/testkit/fixtures/grok_background_subagent/grok_transcript.ndjson

@@ -212,6 +212,18 @@
 {"type":"emit_inbound","label":"notification:_x.ai/session_notification:subagent_progress","frame":{"kind":"notification","method":"_x.ai/session_notification","params":{"sessionId":"00000000-0000-4000-8000-000000000001","update":{"sessionUpdate":"subagent_progress","subagent_id":"00000000-0000-4000-8000-000000000002","attempt_id":"at1.bd80fbaa7c40408eb7999e19714ed4e1","parent_session_id":"00000000-0000-4000-8000-000000000001","child_session_id":"00000000-0000-4000-8000-000000000002","duration_ms":10066,"turn_count":1,"tool_call_count":0,"tokens_used":2912,"context_window_tokens":500000,"context_usage_pct":0,"tools_used":[],"error_count":0}}}}
 {"type":"emit_inbound","label":"notification:_x.ai/task_backgrounded:task_backgrounded","frame":{"kind":"notification","method":"_x.ai/task_backgrounded","params":{"sessionId":"00000000-0000-4000-8000-000000000002","update":{"sessionUpdate":"task_backgrounded","tool_call_id":"call-d5448d70-380d-4a88-a7a0-2968806350ea-0","task_id":"call-d5448d70-380d-4a88-a7a0-2968806350ea-0","command":"sleep 20","cwd":"<workspace>","output_file":"/home/grok-replay/.grok/sessions/%3Cworkspace%3E/00000000-0000-4000-8000-000000000002/terminal/call-d5448d70-380d-4a88-a7a0-2968806350ea-0.log","description":"Sleep for 20 seconds as requested"},"_meta":{"eventId":"00000000-0000-4000-8000-000000000002-170","agentTimestampMs":1790321099801}}}}
 {"type":"emit_inbound","label":"notification:_x.ai/session_notification:background_tasks","frame":{"kind":"notification","method":"_x.ai/session_notification","params":{"sessionId":"00000000-0000-4000-8000-000000000002","update":{"sessionUpdate":"background_tasks","tasks":[{"task_id":"call-d5448d70-380d-4a88-a7a0-2968806350ea-0","command":"sleep 20","description":"Sleep for 20 seconds as requested","cwd":"<workspace>","kind":"bash","status":"running","started_at":"2026-09-25T07:24:44.780647541+00:00","output_file":"/home/grok-replay/.grok/sessions/%3Cworkspace%3E/00000000-0000-4000-8000-000000000002/terminal/call-d5448d70-380d-4a88-a7a0-2968806350ea-0.log"}]},"_meta":{"eventId":"00000000-0000-4000-8000-000000000002-172","agentTimestampMs":1790321099801}}}}
+{"type":"expect_outbound","label":"request:session/prompt:t3-xai-prompt-2","frame":{"kind":"request","method":"session/prompt","params":{"sessionId":"00000000-0000-4000-8000-000000000001","prompt":[{"type":"text","text":"Reply exactly QUEUED_DONE."}],"_meta":{"promptId":"t3-xai-prompt-2","requestId":"t3-xai-prompt-2"}}}}
+{"type":"emit_inbound","label":"notification:session/update:agent_message_chunk:t3-xai-prompt-2","frame":{"kind":"notification","method":"session/update","params":{"sessionId":"00000000-0000-4000-8000-000000000001","update":{"sessionUpdate":"agent_message_chunk","content":{"type":"text","text":"QUEUED"}},"_meta":{"promptId":"t3-xai-prompt-2","agentTimestampMs":1790321085000,"updateType":"AgentMessageChunk","chunkId":1}}}}
+{"type":"emit_inbound","label":"notification:session/update:agent_message_chunk:t3-xai-prompt-2","frame":{"kind":"notification","method":"session/update","params":{"sessionId":"00000000-0000-4000-8000-000000000001","update":{"sessionUpdate":"agent_message_chunk","content":{"type":"text","text":"_DONE"}},"_meta":{"promptId":"t3-xai-prompt-2","agentTimestampMs":1790321085001,"updateType":"AgentMessageChunk","chunkId":2}}}}
+{"type":"emit_inbound","label":"notification:_x.ai/session_notification:turn_completed:t3-xai-prompt-2","frame":{"kind":"notification","method":"_x.ai/session_notification","params":{"sessionId":"00000000-0000-4000-8000-000000000001","update":{"sessionUpdate":"turn_completed","prompt_id":"t3-xai-prompt-2","stop_reason":"end_turn"}}}}
+{"type":"emit_inbound","label":"notification:_x.ai/session/prompt_complete","frame":{"kind":"notification","method":"_x.ai/session/prompt_complete","params":{"sessionId":"00000000-0000-4000-8000-000000000001","promptId":"t3-xai-prompt-2","stopReason":"end_turn","agentResult":null}}}
+{"type":"emit_inbound","label":"response:session/prompt","frame":{"kind":"response","method":"session/prompt","result":{"stopReason":"end_turn","_meta":{"sessionId":"00000000-0000-4000-8000-000000000001","requestId":"t3-xai-prompt-2","promptId":"t3-xai-prompt-2"}}}}
+{"type":"expect_outbound","label":"request:session/prompt:t3-xai-prompt-3","frame":{"kind":"request","method":"session/prompt","params":{"sessionId":"00000000-0000-4000-8000-000000000001","prompt":[{"type":"text","text":"Reply exactly STEER_DONE."}],"_meta":{"promptId":"t3-xai-prompt-3","requestId":"t3-xai-prompt-3"}}}}
+{"type":"emit_inbound","label":"notification:session/update:agent_message_chunk:t3-xai-prompt-3","frame":{"kind":"notification","method":"session/update","params":{"sessionId":"00000000-0000-4000-8000-000000000001","update":{"sessionUpdate":"agent_message_chunk","content":{"type":"text","text":"STEER"}},"_meta":{"promptId":"t3-xai-prompt-3","agentTimestampMs":1790321086000,"updateType":"AgentMessageChunk","chunkId":1}}}}
+{"type":"emit_inbound","label":"notification:session/update:agent_message_chunk:t3-xai-prompt-3","frame":{"kind":"notification","method":"session/update","params":{"sessionId":"00000000-0000-4000-8000-000000000001","update":{"sessionUpdate":"agent_message_chunk","content":{"type":"text","text":"_DONE"}},"_meta":{"promptId":"t3-xai-prompt-3","agentTimestampMs":1790321086001,"updateType":"AgentMessageChunk","chunkId":2}}}}
+{"type":"emit_inbound","label":"notification:_x.ai/session_notification:turn_completed:t3-xai-prompt-3","frame":{"kind":"notification","method":"_x.ai/session_notification","params":{"sessionId":"00000000-0000-4000-8000-000000000001","update":{"sessionUpdate":"turn_completed","prompt_id":"t3-xai-prompt-3","stop_reason":"end_turn"}}}}
+{"type":"emit_inbound","label":"notification:_x.ai/session/prompt_complete","frame":{"kind":"notification","method":"_x.ai/session/prompt_complete","params":{"sessionId":"00000000-0000-4000-8000-000000000001","promptId":"t3-xai-prompt-3","stopReason":"end_turn","agentResult":null}}}
+{"type":"emit_inbound","label":"response:session/prompt","frame":{"kind":"response","method":"session/prompt","result":{"stopReason":"end_turn","_meta":{"sessionId":"00000000-0000-4000-8000-000000000001","requestId":"t3-xai-prompt-3","promptId":"t3-xai-prompt-3"}}}}
 {"type":"emit_inbound","label":"notification:session/update:tool_call_update:01a0d773-bfa9-72a1-b268-790ccebae43f","frame":{"kind":"notification","method":"session/update","params":{"sessionId":"00000000-0000-4000-8000-000000000002","update":{"sessionUpdate":"tool_call_update","toolCallId":"call-d5448d70-380d-4a88-a7a0-2968806350ea-0","status":"completed","title":"[bg] sleep 20 (call-d54)","content":[{"type":"content","content":{"type":"text","text":"Command \"Sleep for 20 seconds as requested\" has been automatically moved to background because it exceeded auto-background timeout limit of 15s. Process is still running."}}],"rawOutput":{"type":"BackgroundTaskStarted","task_id":"call-d5448d70-380d-4a88-a7a0-2968806350ea-0","task_type":"bash","output_file":"/home/grok-replay/.grok/sessions/%3Cworkspace%3E/00000000-0000-4000-8000-000000000002/terminal/call-d5448d70-380d-4a88-a7a0-2968806350ea-0.log","status":"running","command":"sleep 20","summary":"Command \"Sleep for 20 seconds as requested\" has been automatically moved to background because it exceeded auto-background timeout limit of 15s. Process is still running.","retrieval_hint":"Use get_command_or_subagent_output with task_ids=[\"call-d5448d70-380d-4a88-a7a0-2968806350ea-0\"] when you need the output.","pid":2584853}},"_meta":{"totalTokens":13313,"eventId":"00000000-0000-4000-8000-000000000002-173","agentTimestampMs":1790321099802,"promptId":"01a0d773-bfa9-72a1-b268-790ccebae43f","streamStartMs":1790321082753,"turnStartMs":1790321082313,"updateType":"ToolCallUpdate","updateParams":{"toolCallId":"call-d5448d70-380d-4a88-a7a0-2968806350ea-0","status":"Completed"}}}}}
 {"type":"emit_inbound","label":"notification:_x.ai/session_notification:subagent_progress","frame":{"kind":"notification","method":"_x.ai/session_notification","params":{"sessionId":"00000000-0000-4000-8000-000000000001","update":{"sessionUpdate":"subagent_progress","subagent_id":"00000000-0000-4000-8000-000000000002","attempt_id":"at1.bd80fbaa7c40408eb7999e19714ed4e1","parent_session_id":"00000000-0000-4000-8000-000000000001","child_session_id":"00000000-0000-4000-8000-000000000002","duration_ms":18067,"turn_count":1,"tool_call_count":1,"tokens_used":13480,"context_window_tokens":500000,"context_usage_pct":2,"tools_used":["run_terminal_command"],"error_count":0}}}}
 {"type":"emit_inbound","label":"notification:session/update:agent_thought_chunk:01a0d773-bfa9-72a1-b268-790ccebae43f","frame":{"kind":"notification","method":"session/update","params":{"sessionId":"00000000-0000-4000-8000-000000000002","update":{"sessionUpdate":"agent_thought_chunk","content":{"type":"text","text":"The"}},"_meta":{"totalTokens":13480,"eventId":"00000000-0000-4000-8000-000000000002-177","agentTimestampMs":1790321101461,"promptId":"01a0d773-bfa9-72a1-b268-790ccebae43f","streamStartMs":1790321100147,"turnStartMs":1790321082313,"updateType":"AgentThoughtChunk","chunkId":1}}}}

apps/server/src/orchestration-v2/testkit/fixtures/grok_background_subagent/input.ts

@@ -2,19 +2,33 @@ import type { OrchestratorFixtureInput } from "../shared.ts";
 
 export const GROK_BACKGROUND_SUBAGENT_PROMPT =
   "Spawn one background subagent (run it in the background, do not wait for it) whose task is: run the shell command 'sleep 20' and then reply exactly SUBAGENT_DONE. As soon as the subagent is launched, end your turn by replying exactly ROOT_DONE without waiting for it.";
+export const GROK_BACKGROUND_SUBAGENT_QUEUED_PROMPT = "Reply exactly QUEUED_DONE.";
+export const GROK_BACKGROUND_SUBAGENT_STEER_PROMPT = "Reply exactly STEER_DONE.";
 
 const GROK_SUBAGENT_WAKE_LABEL =
   "notification:session/update:agent_message_chunk:subagent-completed-00000000-0000-4000-8000-000000000002";
 
-// The root prompt settles while the subagent still runs, so run 1 is held open
-// until `subagent_finished`, then finishes through the adapter's debounce.
+// The root prompt settles while the subagent still runs. Run 1 completes with
+// its reply and the subagent carries over, so a message queued behind it
+// starts at once (run 2), and a steer aimed at it starts a new turn (run 3)
+// instead of superseding its finished reply. The queued and steered prompts
+// are hand-written in the recorded Grok 1.0.41 shape between the subagent's
+// own frames (#17159). Grok's reply to the finished subagent replays as a
+// continuation run (run 4), like Claude and Codex background wakes.
 export function grokBackgroundSubagentInput(): OrchestratorFixtureInput {
   return {
     steps: [
       { type: "message", text: GROK_BACKGROUND_SUBAGENT_PROMPT },
-      { type: "finish_held_run", targetRunIndex: 1, status: "completed" },
+      { type: "queue_message", text: GROK_BACKGROUND_SUBAGENT_QUEUED_PROMPT },
+      { type: "capture_shell_snapshot", key: "while-subagent-runs" },
+      {
+        type: "steer",
+        text: GROK_BACKGROUND_SUBAGENT_STEER_PROMPT,
+        targetRunIndex: 1,
+        targetReplyFinished: true,
+      },
       { type: "release_replay_gate", label: GROK_SUBAGENT_WAKE_LABEL },
-      { type: "finish_held_run", targetRunIndex: 2, status: "completed" },
+      { type: "finish_held_run", targetRunIndex: 4, status: "completed" },
     ],
   };
 }

apps/server/src/orchestration-v2/testkit/fixtures/grok_background_subagent/output.ts

@@ -9,7 +9,11 @@ import {
   backgroundNotifications,
   projectionFor,
 } from "../shared.ts";
-import { GROK_BACKGROUND_SUBAGENT_PROMPT } from "./input.ts";
+import {
+  GROK_BACKGROUND_SUBAGENT_PROMPT,
+  GROK_BACKGROUND_SUBAGENT_QUEUED_PROMPT,
+  GROK_BACKGROUND_SUBAGENT_STEER_PROMPT,
+} from "./input.ts";
 
 function runAssistantTexts(
   projection: ReturnType<typeof projectionFor>,
@@ -22,21 +26,29 @@ function runAssistantTexts(
 
 // A background subagent outlives its root turn. Grok ends it only with the
 // root-session `subagent_finished` notification: the spawn tool completed at
-// launch and nothing names the subagent done in text.
+// launch and nothing names the subagent done in text. Run 1 completes with its
+// reply while the subagent runs on (#17159): the thread waits on the subagent,
+// a queued message starts at once, and a steer starts a new turn instead of
+// superseding the finished reply. Grok's own reply to the finished subagent
+// (its `subagent-completed-*` wake) is a provider continuation.
 export function assertGrokBackgroundSubagentOutput(
   result: OrchestratorV2ScenarioResult,
   transcript: ProviderReplayTranscript,
 ) {
   assertBaseProjection({
     result,
     transcript,
-    runCount: 2,
-    runStatuses: ["completed", "completed"],
+    runCount: 4,
+    runStatuses: ["completed", "completed", "completed", "completed"],
   });
   const projection = projectionFor(result, transcript.scenario);
   assertSemanticProjectionIntegrity(projection);
-  assertUserMessagesInclude(projection, [GROK_BACKGROUND_SUBAGENT_PROMPT]);
-  const [rootRun, wakeRun] = projection.runs;
+  assertUserMessagesInclude(projection, [
+    GROK_BACKGROUND_SUBAGENT_PROMPT,
+    GROK_BACKGROUND_SUBAGENT_QUEUED_PROMPT,
+    GROK_BACKGROUND_SUBAGENT_STEER_PROMPT,
+  ]);
+  const [rootRun, queuedRun, steerRun, wakeRun] = projection.runs;
 
   assert.lengthOf(projection.subagents, 1);
   const subagent = projection.subagents[0];
@@ -47,6 +59,7 @@ export function assertGrokBackgroundSubagentOutput(
   assert.equal(subagentItem?.status, "completed");
   assert.equal(subagentItem?.runId, rootRun?.id);
 
+  // Run 1 completes with ROOT_DONE; the running subagent does not hold it.
   const subagentCompletedAt = result.domainEvents.findIndex(
     (event) =>
       event.type === "turn-item.updated" &&
@@ -60,24 +73,46 @@ export function assertGrokBackgroundSubagentOutput(
       event.payload.status === "completed",
   );
   assert.isAtLeast(subagentCompletedAt, 0);
-  assert.isAbove(rootCompletedAt, subagentCompletedAt, "run 1 completed before the subagent");
+  assert.isAtLeast(rootCompletedAt, 0);
+  assert.isBelow(rootCompletedAt, subagentCompletedAt, "run 1 waited for the subagent");
   assert.include(runAssistantTexts(projection, rootRun?.id), "ROOT_DONE");
 
-  const rootDoneCompletedAt = result.domainEvents.findIndex(
+  // While the subagent runs, clients see the thread waiting on it, not working.
+  const waitingShell = result.capturedShellSnapshots
+    .get("while-subagent-runs")
+    ?.threads.find((thread) => thread.id === projection.thread.id);
+  assert.deepEqual(
+    waitingShell?.pendingBackgroundTasks?.map((task) => [
+      task.kind,
+      task.kind === "subagent" ? task.childThreadId : undefined,
+    ]),
+    [["subagent", subagent?.childThreadId]],
+    "the Waiting strip names the running subagent",
+  );
+  assert.isNull(waitingShell?.activeRunId, "no run is working while only the subagent runs");
+
+  // The queued message did not wait for the subagent, and the steer started a
+  // new turn: no finished reply was superseded.
+  assert.deepEqual(runAssistantTexts(projection, queuedRun?.id), ["QUEUED_DONE"]);
+  assert.deepEqual(runAssistantTexts(projection, steerRun?.id), ["STEER_DONE"]);
+  const queuedStartedAt = result.domainEvents.findIndex(
     (event) =>
-      event.type === "turn-item.updated" &&
-      event.payload.runId === rootRun?.id &&
-      event.payload.type === "assistant_message" &&
-      event.payload.text.trim() === "ROOT_DONE" &&
-      event.payload.status === "completed",
+      event.type === "run.updated" &&
+      event.payload.id === queuedRun?.id &&
+      event.payload.status === "running",
   );
-  assert.isAtLeast(rootDoneCompletedAt, 0);
+  assert.isAtLeast(queuedStartedAt, 0);
   assert.isBelow(
-    rootDoneCompletedAt,
+    queuedStartedAt,
     subagentCompletedAt,
-    "ROOT_DONE streamed until the subagent finished",
+    "the queued message waited for the subagent",
+  );
+  assert.deepEqual(
+    projection.attempts.map((attempt) => attempt.status),
+    ["completed", "completed", "completed", "completed"],
   );
 
+  // The subagent's own work, done after run 1 settled, lands in its child thread.
   if (subagent?.childThreadId == null) {
     throw new Error("The background subagent is missing its child thread.");
   }
@@ -89,7 +124,16 @@ export function assertGrokBackgroundSubagentOutput(
     ),
     "the subagent's reply must land in its child thread",
   );
-  assert.notInclude(runAssistantTexts(projection, rootRun?.id), "SUBAGENT_DONE");
+  const childTools = child?.turnItems.filter((item) => item.type === "dynamic_tool") ?? [];
+  // Its poll of the sleep, made after run 1 completed, still projects.
+  assert.include(
+    childTools.map((item) => `${item.title}:${item.status}`),
+    "sleep 20 (call-d54):completed",
+    "the subagent's tool calls must land in its child thread",
+  );
+  for (const run of projection.runs) {
+    assert.notInclude(runAssistantTexts(projection, run.id), "SUBAGENT_DONE");
+  }
 
   assert.deepEqual(backgroundNotifications(projection), [
     {