A chat-wake reply is delivered by the system, not by prompt compliance

A wake run whose model answered in plain final text (the coding-tuned
Kimi family does this instead of calling send_message) completed
'successfully', closed its vehicle task, and left the founder staring
at a silent conversation — the reply lived only in provider_requests.
deliverChatReplyIfMissing now runs before a chat-wake run may complete:
when the run never posted into the source thread (run-start watermark,
so a send_message it did make suppresses the fallback), its final text
is delivered as the agent's reply. A delivery failure fails the run
loudly instead of closing it on silence.

tests/integration/ceo-founder-chat.test.ts

@@ -21,6 +21,23 @@ function call(name: string, args: Record<string, unknown>): ToolCall {
   return { id: `call_${randomUUID()}`, name, arguments: args };
 }
 
+/**
+ * Scripts a CEO who never calls send_message: the chat-wake run is answered
+ * with plain final text only — the way a direct-answering model (e.g. a
+ * coding-tuned one) responds when it treats the wake as a question, not a
+ * conversation to post into.
+ */
+function plainTextCeo(): (input: ProviderToolTurnInput) => Promise<AgentTurnResult> {
+  return async (input: ProviderToolTurnInput): Promise<AgentTurnResult> => {
+    const firstUser = input.messages.find((message) => message.role === "user");
+    const prompt = firstUser?.text ?? "";
+    if (!/thread_id="/.test(prompt)) {
+      return { text: "Kickoff noted.", toolCalls: [], rawResponseJson: "{}" };
+    }
+    return { text: "We are on track — the kickoff wave is planned.", toolCalls: [], rawResponseJson: "{}" };
+  };
+}
+
 /**
  * Scripts the CEO: the kickoff run ends quietly; the chat-wake run (its
  * prompt carries the thread transcript) is captured and answered with a
@@ -123,4 +140,77 @@ describe("founder⇄CEO chat", () => {
 
     connection.close();
   });
+
+  test("a CEO who answers in plain text still has the reply delivered into the founder's thread", async () => {
+    const connection = new DatabaseConnection(join(tmpdir(), `ac-ceo-plain-${randomUUID()}.db`));
+    const providers = new ProviderService(connection.db, providerAdapters, testVault, undefined, plainTextCeo());
+    const onboarding = new OnboardingService({ db: connection.db, saveProviderCredential: (input) => providers.saveCredential(input) });
+    const taskExecution = new TaskExecutionService(connection.db, providers);
+    const chat = new ChatService(connection.db);
+
+    let orgId = "";
+    let ceoId = "";
+    connection.transaction(() => {
+      const summary = onboarding.bootstrap({
+        companyName: "FlowPilot",
+        companyDescription: "CSM analytics for SMBs.",
+        employeeName: "Yang",
+        workspacePath: "",
+        executionMode: "ai_ceo",
+        autoApproveHires: true,
+        reviewDeliverables: false,
+        goalTitle: "Ship the MVP.",
+        goalDescription: "A working product.",
+        providerKey: "openai",
+        providerSecret: "sk-test-key",
+        providerModel: "gpt-5.4",
+      });
+      orgId = summary.orgId;
+      ceoId = summary.ceoAgentId;
+    });
+    const founder = connection.db.prepare("select id from employees where org_id = ?").get(orgId) as { id: string };
+    const pinnedSpaceId = ensureCeoDirectSpace(connection.db, orgId);
+
+    // Drain the kickoff run so the next queued run is the chat wake.
+    const kickoff = claimNextQueuedRun(connection);
+    await taskExecution.executeRun(kickoff!.org_id, kickoff!.id);
+
+    const founderMessageId = chat.sendMessage({
+      orgId,
+      spaceId: pinnedSpaceId!,
+      threadId: null,
+      senderType: "employee",
+      senderId: founder.id,
+      bodyMd: "How are we tracking against the MVP?",
+      attachmentFileIds: [],
+    });
+    const founderThreadId = (connection.db.prepare("select thread_id from chat_messages where id = ?")
+      .get(founderMessageId) as { thread_id: string }).thread_id;
+
+    const wake = connection.db.prepare(
+      "select id from task_runs where org_id = ? and agent_id = ? and trigger = 'chat_mention' and state = 'queued'",
+    ).get(orgId, ceoId) as { id: string };
+    await taskExecution.executeRun(orgId, wake.id);
+
+    // The model answered in its final text and never called send_message; the
+    // reply must still land in the founder's thread — exactly once.
+    const replies = connection.db.prepare(
+      "select space_id, thread_id, body_md from chat_messages where org_id = ? and sender_type = 'agent' and sender_id = ?",
+    ).all(orgId, ceoId) as Array<{ space_id: string; thread_id: string; body_md: string }>;
+    expect(replies).toHaveLength(1);
+    expect(replies[0].space_id).toBe(pinnedSpaceId);
+    expect(replies[0].thread_id).toBe(founderThreadId);
+    expect(replies[0].body_md).toContain("on track");
+
+    // The run completed and closed its wake task, not failed on a delivery it made.
+    const wakeRun = connection.db.prepare("select state, error_message from task_runs where id = ?")
+      .get(wake.id) as { state: string; error_message: string | null };
+    expect(wakeRun).toMatchObject({ state: "completed", error_message: null });
+    const wakeTask = connection.db.prepare(
+      "select t.status from tasks t join task_runs r on r.task_id = t.id where r.id = ?",
+    ).get(wake.id) as { status: string };
+    expect(wakeTask).toEqual({ status: "done" });
+
+    connection.close();
+  });
 });