FN-7765: fix dashboard chat crash when plugin CLI sessions omit session.state

Fix a "Response failed" crash for plugin CLI runtime chats (grok/droid/cursor) whose sessions expose top-level `messages` and stream via `onText` without a pi-shaped `session.state`.

- Read messages/errorMessage null-safely from `session.state`, falling back to top-level `session.messages` when state is absent, in both the room responder and streaming response extraction paths
- Keep `state.errorMessage` optional so successful streams from state-less sessions no longer throw TypeErrors, while pi/openclaw/hermes provider errors still surface correctly
- Add regression tests covering state-less plugin CLI sessions in chat-manager.test.ts
- Add changeset for the fix

Files changed:
 .changeset/fn-7765-grok-cli-chat-crash.md          |  7 ++
 packages/dashboard/src/__tests__/chat-manager.test.ts | 90 ++++++++++++++++++++++
 packages/dashboard/src/chat.ts                     | 35 ++++++---
 3 files changed, 121 insertions(+), 11 deletions(-)

Fusion-Task-Id: FN-7765

Fusion-Task-Lineage: 80e33a97-e971-4aec-a4cf-29d97c5c5e62

Co-authored-by: Fusion (runfusion.ai) <noreply@runfusion.ai>

.changeset/fn-7765-grok-cli-chat-crash.md

@@ -0,0 +1,7 @@
+---
+"@runfusion/fusion": patch
+---
+
+summary: Fix Grok CLI chat failing instantly with a "Response failed" error.
+category: fix
+dev: ChatManager.sendMessage (packages/dashboard/src/chat.ts) now null-safely reads session.state.errorMessage/messages and falls back to the session's top-level messages + accumulated onText stream, so plugin-backed CLI runtime sessions (grok/droid/cursor) that expose no pi-shaped `state` render their reply instead of throwing "Cannot read properties of undefined (reading 'errorMessage')". pi/openclaw/hermes state.errorMessage failure bubbles are unchanged. Same fix applied to the room-responder session.state.messages read.

packages/dashboard/src/__tests__/chat-manager.test.ts

@@ -547,6 +547,96 @@ describe("ChatManager.sendMessage", () => {
     expect(assistantCall?.[1].content).toBe("Hello world!");
   });
 
+  it("persists streamed replies and broadcasts done for no-state plugin runtime sessions", async () => {
+    const events: Array<{ type: string; data: any }> = [];
+    const unsubscribe = chatStreamManager.subscribe("chat-001", (event) => {
+      events.push(event);
+    });
+    mockChatStore.addMessage.mockImplementation((_sessionId, input) => ({
+      id: input.role === "assistant" ? "assistant-msg" : "user-msg",
+      sessionId: "chat-001",
+      role: input.role,
+      content: input.content,
+      thinkingOutput: input.thinkingOutput,
+      metadata: input.metadata,
+      createdAt: "2026-07-10T00:00:00.000Z",
+    }));
+
+    __setCreateFnAgent(async (options: any) => ({
+      session: {
+        prompt: vi.fn().mockImplementation(async () => {
+          options.onText?.("Grok CLI streamed reply");
+        }),
+        dispose: vi.fn(),
+        messages: [],
+      },
+    }));
+
+    const chatManager = createChatManager();
+    await expect(chatManager.sendMessage("chat-001", "Hello Grok")).resolves.toBeUndefined();
+    unsubscribe();
+
+    const assistantCalls = mockChatStore.addMessage.mock.calls.filter((call) => call[1].role === "assistant");
+    expect(assistantCalls).toHaveLength(1);
+    expect(assistantCalls[0]?.[1].content).toBe("Grok CLI streamed reply");
+    expect(events).toContainEqual(expect.objectContaining({
+      type: "done",
+      data: expect.objectContaining({
+        message: expect.objectContaining({ content: "Grok CLI streamed reply" }),
+      }),
+    }));
+    expect(events).not.toContainEqual(expect.objectContaining({ type: "error" }));
+    expect(assistantCalls[0]?.[1].content).not.toContain("Response failed");
+  });
+
+  it("does not crash when a no-state plugin runtime streams no text", async () => {
+    const events: Array<{ type: string; data: any }> = [];
+    const unsubscribe = chatStreamManager.subscribe("chat-001", (event) => {
+      events.push(event);
+    });
+    mockChatStore.addMessage.mockImplementation((_sessionId, input) => ({
+      id: input.role === "assistant" ? "assistant-msg" : "user-msg",
+      sessionId: "chat-001",
+      role: input.role,
+      content: input.content,
+      createdAt: "2026-07-10T00:00:00.000Z",
+    }));
+
+    __setCreateFnAgent(async () => ({
+      session: {
+        prompt: vi.fn().mockResolvedValue(undefined),
+        dispose: vi.fn(),
+        messages: [],
+      },
+    }));
+
+    const chatManager = createChatManager();
+    await expect(chatManager.sendMessage("chat-001", "Hello Grok")).resolves.toBeUndefined();
+    unsubscribe();
+
+    const assistantCalls = mockChatStore.addMessage.mock.calls.filter((call) => call[1].role === "assistant");
+    expect(assistantCalls).toHaveLength(1);
+    expect(assistantCalls[0]?.[1].content).toBe("");
+    expect(events).toContainEqual(expect.objectContaining({ type: "done" }));
+    expect(events).not.toContainEqual(expect.objectContaining({ type: "error" }));
+  });
+
+  it("falls back to top-level messages for no-state plugin runtime sessions", async () => {
+    __setCreateFnAgent(async () => ({
+      session: {
+        prompt: vi.fn().mockResolvedValue(undefined),
+        dispose: vi.fn(),
+        messages: [{ role: "assistant", content: "Top-level Grok message" }],
+      },
+    }));
+
+    const chatManager = createChatManager();
+    await expect(chatManager.sendMessage("chat-001", "Hello Grok")).resolves.toBeUndefined();
+
+    const assistantCall = mockChatStore.addMessage.mock.calls.find((call) => call[1].role === "assistant");
+    expect(assistantCall?.[1].content).toBe("Top-level Grok message");
+  });
+
   // U11 / R12 drift guard: the chat lane must expose workflow discovery,
   // mutation, settings, selection, and trait vocabulary to the agent when a
   // scoped task store is available.

packages/dashboard/src/chat.ts

@@ -1781,7 +1781,13 @@ export class ChatManager {
       );
 
       type AgentMessage = { role?: string; type?: string; content?: string | Array<{ type?: string; text?: string }> };
-      const messages = (resolvedSession.session.state.messages as AgentMessage[]) ?? [];
+      /*
+       * FNXC:Chat 2026-07-10-00:00:
+       * Plugin CLI runtime sessions (grok/droid/cursor) expose top-level `messages` and stream via `onText` without a pi-shaped `state`, so room responders must read messages null-safely while preserving pi/openclaw state-backed sessions.
+       */
+      const roomSessionState = resolvedSession.session.state as { messages?: AgentMessage[]; errorMessage?: string } | undefined;
+      const roomTopLevelMessages = (resolvedSession.session as { messages?: AgentMessage[] }).messages;
+      const messages = roomSessionState?.messages ?? roomTopLevelMessages ?? [];
       const lastAssistant = [...messages].reverse().find((message) => message.role === "assistant" || message.type === "assistant");
       let content = "";
       if (typeof lastAssistant?.content === "string") {
@@ -1792,7 +1798,7 @@ export class ChatManager {
           .join("");
       }
 
-      const stateError = (resolvedSession.session.state as { errorMessage?: string } | undefined)?.errorMessage;
+      const stateError = roomSessionState?.errorMessage;
       if (stateError?.trim()) {
         throw new Error(stateError.trim());
       }
@@ -2369,10 +2375,16 @@ export class ChatManager {
         return;
       }
 
-      // Some runtimes (e.g. plugin-backed Codex/openclaw) signal provider failures
-      // by setting session.state.errorMessage rather than throwing. Surface that
-      // as an error event instead of persisting a blank assistant reply.
-      const sessionErrorMessage = (agentResult.session.state as { errorMessage?: unknown }).errorMessage;
+      interface AgentMessage {
+        role: string;
+        content?: string | Array<{ type: string; text: string }>;
+      }
+      /*
+       * FNXC:Chat 2026-07-10-00:00:
+       * Plugin CLI runtime sessions (grok/droid/cursor) expose top-level `messages` and stream via `onText` without a pi-shaped `state`; keep `state.errorMessage` optional so successful streams do not become TypeErrors, while pi/openclaw/hermes provider errors still surface when set.
+       */
+      const agentSessionState = agentResult.session.state as { errorMessage?: unknown; messages?: AgentMessage[] } | undefined;
+      const sessionErrorMessage = agentSessionState?.errorMessage;
       if (typeof sessionErrorMessage === "string" && sessionErrorMessage.trim().length > 0
           && !accumulatedText && !accumulatedThinking && toolCallsAccum.length === 0) {
         const failureInfo = addModelContextToFailureInfo(
@@ -2391,11 +2403,12 @@ export class ChatManager {
 
       // Extract response text from agent state
       let responseText = "";
-      interface AgentMessage {
-        role: string;
-        content?: string | Array<{ type: string; text: string }>;
-      }
-      const lastMessage = (agentResult.session.state.messages as AgentMessage[])
+      /*
+       * FNXC:Chat 2026-07-10-00:00:
+       * Plugin CLI runtimes can omit `state` entirely; use streamed text first, then fall back through state-backed messages and top-level session messages so no-state sessions persist successful replies instead of crashing during extraction.
+       */
+      const agentMessages = agentSessionState?.messages ?? (agentResult.session as { messages?: AgentMessage[] }).messages ?? [];
+      const lastMessage = agentMessages
         .filter((m: AgentMessage) => m.role === "assistant")
         .pop();