Design agent appears frozen during expert consultation — no progress indicator for ~2 minutes

Bug: Design Agent Appears Frozen During Expert Consultation Phase

Summary

When the Tier 2 design agent enters the expert consultation phase (parallel Gemini Flash reviews), the UI shows no progress updates for approximately 2 minutes. The agent IS working (backend processing 8 parallel expert reviews), but the frontend only shows a static "Still working..." text with a timer. No new tool call cards, thinking blocks, or status updates appear during this period, making it look like the system has crashed or frozen.

Steps to Reproduce

1. Sign into MANTI at localhost:8081 or mantiai.com
2. Start a new chat and send this prompt:

``
/design I want to build a full slam robot with 3 main components. A esp32-s3 Arduino powered robot car with an LD19 lidar, a single accelerometer, and 2 encoders on each wheel. The ESP32 must communicate over wifi to the main server running on my personal computer, most likely a python server that controls and receives the information from the robot. Then we need to help the user visualize and see what is going on so lets build it as a frontend server in react that communicates with the backend server and allows us to select locations for the robot to go to both on the map or even in exploration mode to find what is happening. NO ROS
``

1. Click "Yes, switch to advanced" when the mode switch prompt appears
2. Answer all 5 clarifying questions (Budget: $100-200, Environment: Apartment, Timeline: 1 month, Experience: Intermediate, Purpose: Learning project)
3. Click "Submit All Answers"
4. Observe: After submitting answers, the agent processes them and enters the "Consulting expert panel" phase

What Happens

• The agent shows "Consulting expert panel" as a collapsed tool call card
• Below it shows "Still working..." with a running timer (e.g., "3m 52s", "4m 7s")
• For approximately 2 minutes, nothing else updates on screen
• No new thinking blocks appear
• No new tool call cards appear
• No streaming text appears
• The "Stop generation" button is present (indicating the agent IS running)
• The send button is disabled (correct — agent is processing)

What It Looks Like to a User

• The system appears to have frozen or crashed
• Users will likely try to reload the page, which can disrupt the agent run
• There's no way to distinguish "processing" from "stuck"
• After the ~2 minute silent period, the agent suddenly resumes with new document sections

Root Cause

The expert consultation sub-agent (consultExperts()) runs 8 parallel Gemini Flash calls internally. During this time:

• The BullMQ job is active (not stalled)
• The agent loop is blocked waiting for the expert consultation tool result
• No intermediate events are emitted to the frontend WebSocket
• The only visible indicator is the "Still working..." static text

Expected Behavior

During expert consultation, the UI should show intermediate progress:

• "Reviewing with embedded systems expert... (1/8)"
• "Reviewing with SLAM algorithm specialist... (3/8)"
• "Expert panel review complete — synthesizing feedback"

Or at minimum, a pulsing/animated indicator that clearly shows active processing vs frozen state.

Environment

• chat-service with Claude Sonnet 4.6 (Tier 2)
• Expert consultation uses Gemini Flash (8 parallel reviews)
• Observed during local E2E testing (March 2026)

Impact

• User Experience: Users think the system is broken and reload/abandon
• Trust: Undermines confidence in the platform during the most impressive feature (autonomous design)
• Priority: Week 2 polish — not a functional blocker but critical for user perception

Suggested Fix

Emit progress events from expert-debate.sub-agent.ts as each expert review completes. The events should flow through the existing Redis pub/sub → WebSocket pipeline and render as a progress indicator in the expert panel tool card.

Key files:

• services/chat-service/src/tiers/tier2-advanced/experts/expert-debate.sub-agent.ts
• frontend/src/components/chat/cards/ (tool card rendering)