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.

src/main/services/chat-agent-wake-service.ts

@@ -4,6 +4,7 @@ import { RISK_LEVEL, TASK_PRIORITY, TASK_STATUS } from "../../shared/domain";
 import { createId } from "../../shared/ids";
 import type { ResolvedMention } from "./chat-mentions";
 import type { SlashInvocation } from "./chat-service-internal";
+import type { ChatService } from "./chat-service";
 import { OrganizationService } from "./organization-service";
 
 interface RunnableAgentRow {
@@ -63,6 +64,51 @@ export function expireStaleChatWakes(db: DatabaseSync, now: string): number {
   return stale.length;
 }
 
+/**
+ * A chat wake's deliverable is a reply VISIBLE in the source thread. The model
+ * is asked to post it with send_message, but a run that answers in plain final
+ * text has still produced the reply — delivery is the system's contract, not
+ * the model's prompt compliance. When the run never posted into the thread
+ * (checked by the run's start watermark, so any send_message it did make
+ * counts), its final text is delivered as the agent's message; otherwise the
+ * asker watches the run "complete" while the conversation stays silent.
+ * Returns true when a fallback message was posted. A delivery failure throws,
+ * so the run fails loudly instead of completing with nothing said.
+ */
+export function deliverChatReplyIfMissing(
+  db: DatabaseSync,
+  chat: ChatService,
+  input: {
+    orgId: string;
+    agentId: string;
+    spaceId: string;
+    threadId: string;
+    finalText: string;
+    /** ISO timestamp captured before the run's first provider turn. */
+    runStartedAt: string;
+  },
+): boolean {
+  const body = input.finalText.trim();
+  if (!body) return false;
+  const posted = db.prepare(`
+    select 1 from chat_messages
+    where org_id = ? and thread_id = ? and sender_type = 'agent' and sender_id = ? and created_at >= ?
+    limit 1
+  `).get(input.orgId, input.threadId, input.agentId, input.runStartedAt);
+  if (posted) return false;
+  console.info(`[chat] the reply run never posted into thread ${input.threadId}; delivering its final text as the reply.`);
+  chat.sendMessage({
+    orgId: input.orgId,
+    spaceId: input.spaceId,
+    threadId: input.threadId,
+    senderType: "agent",
+    senderId: input.agentId,
+    bodyMd: body,
+    attachmentFileIds: [],
+  });
+  return true;
+}
+
 export class ChatAgentWakeService {
   private readonly organization: OrganizationService;
 

src/main/services/task-execution-service.ts

@@ -7,6 +7,7 @@ import { AuditService } from "./audit-service";
 import { ProviderService } from "./provider-service";
 import { OrganizationService } from "./organization-service";
 import { ChatService } from "./chat-service";
+import { deliverChatReplyIfMissing } from "./chat-agent-wake-service";
 import { SkillService } from "./skill-service";
 import { runAgentLoop } from "./agent-run-loop";
 import type { AgentToolContext } from "./agent-tools";
@@ -151,6 +152,9 @@ export class TaskExecutionService {
         .map(connectorDisplayName);
       const language = this.appLanguage();
       const externalTools = await loadComposioTools(this.db, orgId, agent.id);
+      // Watermark for the reply-delivery check below: anything the agent posts
+      // into the source thread from here on counts as its reply.
+      const runStartedAt = new Date().toISOString();
       const result = await runAgentLoop(
         {
           providers: this.providers,
@@ -183,6 +187,21 @@ export class TaskExecutionService {
         },
       );
       providerRequestId = result.lastRequestId;
+      // A chat wake's deliverable is the visible reply. If the model answered
+      // in plain text instead of calling send_message, deliver that text into
+      // the source thread BEFORE the run may complete — a delivery failure
+      // fails the run rather than closing it on silence.
+      if (chatContext) {
+        const delivered = deliverChatReplyIfMissing(this.db, this.chat, {
+          orgId,
+          agentId: agent.id,
+          spaceId: chatContext.spaceId,
+          threadId: chatContext.threadId,
+          finalText: result.finalText,
+          runStartedAt,
+        });
+        if (delivered) this.notifyDataChanged();
+      }
       this.completeRun(orgId, runId, result.lastRequestId);
       this.clearProviderCooldown(provider);
       this.audit.record({
@@ -511,6 +530,9 @@ export class TaskExecutionService {
       return null;
     }
     if (!scope.sourceThreadId) return null;
+    const thread = this.db.prepare("select space_id from chat_threads where id = ? and org_id = ?")
+      .get(scope.sourceThreadId, orgId) as { space_id: string } | undefined;
+    if (!thread) return null;
     const messages = this.db.prepare(`
       select m.body_md, m.sender_type, m.sender_id,
              coalesce(a.name, e.display_name, m.sender_type) as sender_name
@@ -528,6 +550,7 @@ export class TaskExecutionService {
       m.sender_type === "employee" || m.sender_type === "board" ? `${m.sender_name} (founder)` : m.sender_name;
     return {
       threadId: scope.sourceThreadId,
+      spaceId: thread.space_id,
       transcript: messages.map((m) => `[${label(m)}]: ${m.body_md}`).join("\n"),
     };
   }

src/main/services/task-prompts.ts

@@ -68,6 +68,8 @@ export interface SkillRuntimeSnapshotRow {
 
 export interface ChatContextBlock {
   threadId: string;
+  /** The space the thread lives in — where the run's reply must land. */
+  spaceId: string;
   transcript: string;
 }