Allow long-running ACP requests to complete asynchronously over SSE

[RGHenderson]

Problem

ACP requests such as session/prompt can run for several minutes. The HTTP adapter currently keeps the POST open until the agent returns its JSON-RPC response. This makes correctness depend on every proxy and client allowing an arbitrarily long response-header timeout; for example, Node/Undici aborts after its default five-minute headers timeout.

The client already maintains an SSE connection capable of receiving JSON-RPC responses. Holding the originating POST open duplicates that response path and makes a transport timeout capable of disrupting an otherwise healthy stream.

Proposed contract

For requests that will complete asynchronously:

1. Accept and validate the JSON-RPC request.
2. Return 202 Accepted promptly from the POST.
3. Deliver the eventual JSON-RPC result or error on the existing SSE stream, correlated by request ID.
4. Keep immediate 4xx/5xx responses for failures that occur before acceptance.

Notifications can continue to return 202 without a later response.

Questions

• Should this apply to every ACP request, or only explicitly long-running methods such as session/prompt?
• Does ACP-over-streamable-HTTP already define this behavior, or should sandbox-agent document an extension?
• What retry/idempotency guarantees should clients use if the POST connection fails before receiving 202?

Separately, a narrow client reliability fix can ensure an individual detached POST failure rejects only its matching JSON-RPC request rather than closing the shared SSE transport. That protects current behavior but does not remove the long-lived POST dependency described here.

[linear]

RVT-6354