Experiment: ai-elements surfaces

Status: implemented (2026-08). Hypar consumes ai-elements-nuxt 1.5 as a headless UI layer. The agent contract stays JSON { question, connections, … } — not a chatbot UIMessage stream.

The library was already a module in nuxt.config.ts with a terminal theme for [data-ai-*] hooks. No page used an Ai* component. This experiment wires primitives onto the embryo collaborator, the connection graph, the garden queue, and settings.

---

Problem

Hypar's pitch is that surviving interaction patterns are extracted into ai-elements-nuxt. Until this experiment, the extraction pipeline was reversed: the library existed, the lab did not use it. The agent panel was custom HTML; SSE chunk events were ignored; connections were a flat list; settings was a stub.

Thesis

Headless AI UI can serve a collaborator (one question, HITL links, typed graph) without turning the product into a chatbot — if we map domain objects onto primitives instead of adopting useAiChat.

What we mounted

Surface	Components	Domain mapping
agent.collaborate	AiMessage, AiMarkdown, AiPromptInput, AiShimmer, AiStreamingCursor, AiErrorBoundary, AiConfirmation, AiContext, AiSuggestion, AiSpeechInput	Turns, reply, stream preview of question, pending connection/path/fossil
Connections	AiCanvas, AiNode, AiEdge, AiControls	Embryos as nodes; type as edge label; inferred+unconfirmed = animated/dashed
Garden	AiQueue, AiSpeechInput	Unanswered PENDING_QUESTION notes; voice seed capture
Settings	AiModelSelector	Ollama tags (cookie override, sent on the next agent POST)

What we did not mount

useAiChat / createChatHandler, AiSources, AiPlan / AiTask, visible AiReasoning / AiChainOfThought, AiConversation as the garden, coding-agent surfaces (Sandbox, Commit, …). Those either contradict The Agent or belong to later experiments. Method moves are now lightly visible on the challenge hero (from payload.move), without a Munari stepper.

Observations to collect

1. Does streaming the partial question field feel like thinking-with, or like watching JSON form?
2. Do dashed inferred edges make agent speculation visually weaker than user-drawn links? (open question)
3. Does AiContext (seed, tensions, turn count) answer “what does the agent know?” without dumping the whole garden?
4. Which of these survive extraction back into the library (likely: typed inferred-vs-explicit edges on a knowledge graph, not another chat demo)?

Non-goals

• Replacing the embryo event log with AI SDK message persistence.
• Auto-layout beyond a circle around the current embryo.
• Server-side persistence of the selected model (cookie + request body only).