finding: DataBlock is refused on input, so structured context has no typed form

[sagi5060]

What

DataBlock is the content block that exists to carry JSON. The openai-agents engine refuses it
as input:

ConfigError: openai-agents engine cannot send a 'data' block to the model;
it accepts text, image, and audio (chat-completions only) input blocks

Reproduced at ca03ae8 (v3.0.0):

payload = [DataBlock(data={"page": "reference/deck"}), TextBlock(text="what page am I on?")]
await deck.run("AskAgentDeck", payload, context=corpus)   # raises

Raised from adapters/engines/openai_agents/engine.py:271 (_part_of). ResourceBlock takes
the same path.

Why it matters

DataBlock is an output block in practice — the engine produces one for structured output and
refuses one on input — but nothing says so. The consequence showed up building the v3 reference
app (#219): the natural, typed way to hand a model structured per-run context does not exist, so
every embedded application invents its own prose preamble. Ask AgentDeck's is
<context>…</context> in examples/ask-agentdeck/ask_agentdeck/server.py. No two applications
will agree on a format, and each one re-solves delimiter injection from scratch.

reference/deck.mdx lists all five block types under What input accepts and qualifies it a
paragraph later. Accurate, but it reads as an afterthought for something that is closer to a
type error.

Two directions, neither chosen here

• The engine renders a DataBlock as JSON text rather than refusing it — cheap, and arguably
  what a caller means.
• The block table states plainly which types are input-capable, and the docs stop implying
  symmetry the content model does not have.

Version anchor

dev at ca03ae8, v3.0.0. Recorded in docs/delivery/plan-219-delivery.md §4.

What is not yet known

Whether the LangGraph engine behaves the same way — the reproduction above is openai-agents
only, and a divergence between the two engines on the same block type would be its own finding.

[sagi5060]

Related to #227 — same boundary, independent fixes

Both came out of the v3 reference application's friction ledger (plan-219-delivery.md §4) and both are about getting per-run application data to a served agent:

• deck: asgi() cannot serve an agent that needs a context — is the packaged surface demo-only? #227 — a live Python object cannot cross HTTP, so asgi() cannot serve a context-using app.
• This issue — the typed way to send data (not an object) is DataBlock, and the engine refuses it on input.

Together they are why Ask AgentDeck invents a prose <context> preamble. Independently fixable, and worth deciding together: if #227 grows a context factory, the pressure on this one drops; if it does not, this becomes the only structured channel there is. Tracked in docs/delivery/findings-register.md.