The Constellation — a context graph of the harness
Every entity as a node, clustered by type, connected where they relate. Flip a node into context (◆) and it joins a Context Stack — a shareable, agent-consumable overlay. A lens re-lays the whole field by outcome / role / person. This doubles as the harness's own infographic: what it holds and how it feeds itself.
LENS · TYPEOutcomeJob rolePeopleFreshness
Apps
Content
Labs
Showcase
Signal core
◆ CONTEXT STACK · "Board demo"
Snowscape · work
AI Benchmarks · proof
RAI Report · signal
AgentLand · build
FEED TO AGENT ▸
◆ = in context. The stack becomes the agent's working set — Gia sees exactly these when you talk. Like the share Stack, but for thinking, not sending.
What you're looking at
Clusters = types (apps/content/labs/showcase) pulled toward their region; node size + brightness = depth (the "3D Minority Report" tilt — near nodes big/sharp, far ones small/dim). Lines = real relationships (shared focus tags, feeds — e.g. RAI Report → Week in AI → Today in AI form a visible spine). Toggling a node ◆ lifts it into the Context Stack.
⚖ CRITIQUE — is this a toy or the actual thing?
✓ Why it's realYou already have the graph's data — types, focus tags, provenance, feeds between Signal cards. This isn't inventing structure, it's revealing structure you built. And "Context Stack" is a genuinely novel primitive: the share Stack sends cards to a person; a context stack feeds cards to an agent as its working set. Same gesture, opposite direction — outward vs inward.
△ The traps① A pretty graph that doesn't do anything is a screensaver. The payoff has to be the FEED-TO-AGENT edge: context stack → Direct Line working set, or it's decoration. ② Force-directed graphs get hairball-ugly fast past ~40 nodes — you need the lens to always be a hard filter, never "show everything." ③ 3D-for-its-own-sake fights usability on a phone; keep the depth as a subtle tilt, not a camera you fly.
On the name"Context overlay" is close but overloaded. I'd call the primitive a Lens (the view) + a Context Stack (the working set it produces). "Modules" undersells it. The lens re-projects; the stack is what you hand an agent. Two words, two jobs — don't merge them.
💡 The realization worth keeping: this graph IS the harness infographic. One artifact answers "what has Byron built" (the constellation), "how does it connect" (the edges), and "how does an agent consume it" (context stack → feed). It's the Building-Blocks palette idea and the Signal-primitive thesis, drawn as one living picture — and it's the single most demo-able object in the whole system for an RAI-leader audience: "here is my operation, and here is me handing my agent exactly the context I choose."
Status: mockup only. Real version = a Phase ⑤/⑥ once Direct Line exists (the FEED-TO-AGENT edge needs the chat relay to land in). Cheap first step whenever you want it: a static constellation view built from the current registry, no context-feed yet — just to see the patterns your real data forms.