Phase 5a Dashboard — Demo Screencast

Demo recording for the Phase 5a read-only web dashboard.

This is not a code release — it's a media drop. The dashboard code
itself is in commit 80cc8b3
on main.

What you see in the demo

A multi-agent task chain running across two real worker daemons:

• worker-gpt-oss-cloud — gpt-oss:120b via Ollama Cloud (OpenAI OSS family)
• worker-claude-sonnet — claude-sonnet-4-6 via claude -p (Anthropic Max subscription)

The lead (Claude Opus on the Mac) drives the dispatch from a /lead-mode
session, and the dashboard (served by the broker) live-renders the chain
in 1 s polling intervals.

Highlights

• Pre-staged DAG — full task graph drawn in gray ("blocked / pending")
  before any work starts. Sentinel hold pattern lets the user see the
  whole structure before pressing go.
• Cross-vendor parallel execution — heavy reasoning (gpt-oss) and
  heavy prose (claude-sonnet) run side-by-side, durations ticking in
  real time on each box.
• Per-box metadata — agent name, model identifier, and live duration
  shown on every node.
• Connected-component batching — each /lead-mode dispatch becomes
  its own graph; older runs auto-clear when a new chain arrives.
• Dark / cyan tech aesthetic — gray (waiting) → amber (active) →
  green (done) / red (failed) lifecycle, easy to read at a glance.

Code pointers

• Backend: src/mab/broker/routes/dashboard.py
• Frontend: src/mab/broker/static/index.html
• Cookbook of working multi-agent recipes: docs/cookbook.md
• Phase 5a TARGET section: TARGET.md