NNRP/1-preview4 Design: Runtime Object and Control Protocol

NNRP/1-preview3 proved that SDKs can share Rust-backed protocol artifacts and publish consistent preview packages. It did not prove that every workload benefits from replacing HTTP or SSE. The preview4 design therefore moves the center of gravity from transport replacement to runtime object efficiency: keep large or fast-changing runtime state out of repeatedly parsed JSON, make scheduling decisions explicit, and let conformance verify protocol semantics at the wire level.

Positioning

Preview4 is for workloads where the protocol boundary carries meaningful runtime state:

• AI coding agents that schedule subagents, tool calls, artifact streams, and multimodal context.
• Neural rendering or runtime services that need partial results, deadlines, drop reasons, and trace timing instead of all-or-nothing responses.
• Adapter ecosystems where inference engines should not block NNRP adoption even when the model compute path dominates total latency.

It is not a claim that token streaming, vLLM chat completions, or cache lookups are automatically faster than mature HTTP paths. Cache is one lifecycle tool, not the central promise.

Architecture

Preview4 keeps the coarse FFI and runtime boundaries introduced in preview3. The new work adds semantics above those boundaries, not more per-field boundary calls.

Runtime Object Lifecycle

Heavy payloads should be named, referenced, patched, and released. A token-oriented JSON profile is still valid for small responses, but it must not be the only representation for tensors, image regions, tool artifacts, or reusable context.

Control Frame Families

Family	Purpose
CANCEL / ABORT	Stop obsolete work. CANCEL is cooperative; ABORT allows hard teardown when the operation is no longer useful.
SUPERSEDE	Replace an older operation with a newer one while preserving trace continuity.
PRIORITY_UPDATE	Change priority after admission so urgent work can move ahead of stale work.
DEADLINE / EXPIRE_AT	Declare when work becomes useless and may be dropped.
PROGRESS / PARTIAL_RESULT	Return usable progress and partial artifacts before the final result.
BACKPRESSURE / CREDIT_UPDATE	Let either side shrink or expand send windows without inventing an out-of-band throttle.
CAPABILITY_NEGOTIATION	Include support, cost, preference, limits, and downgrade behavior, not only boolean flags.
ROUTE_HINT / EXECUTION_HINT	Carry scheduling hints for subagents and runtime dispatch without heavy JSON or protobuf wrappers.
CACHE_REFERENCE / CACHE_MISS / CACHE_INVALIDATE	Formalize cache references and invalidation where reuse is real, without assuming every scene or prompt is cache-friendly.
TRACE_CONTEXT	Carry end-to-end timing and correlation IDs across frames.
RESULT_DROP_REASON	Explain why a result was dropped, such as deadline expiry, supersession, backpressure, or peer cancellation.
DEGRADE_PROFILE	Negotiate a cheaper profile when the preferred one is not available or too expensive.
BUDGET_UPDATE	Update compute, token, memory, or bandwidth budgets during a session.
ERROR_RECOVERABLE / RETRY_AFTER	Distinguish retryable pressure from terminal failure.

Profile Expansion

Preview4 should add profile families before forcing every workload through the token stream profile:

• runtime.control: common control frames, deadlines, priority, progress, backpressure, trace, and drop reasons.
• runtime.object: object declaration, object references, deltas, release, and cost metadata.
• coding.agent: subagent routing, tool-call artifacts, execution hints, and cancellation policy.
• multimodal.artifact: image/audio/video/tool artifacts with typed descriptors and partial result channels.
• render.runtime: frame deadlines, partial region results, result drop reasons, and trace stages.
• cache.reference: cache reference, miss, invalidation, and lease semantics for workloads that can actually reuse state.

Wire-level Conformance

Preview4 conformance must include an active wire-level runner. SDK adapter tests are still useful, but they can hide semantic drift because the implementation controls both the adapter and local runtime calls.

The wire runner can:

• act as an NNRP client and connect to an implementation server;
• act as an NNRP server and accept an implementation client;
• act as a proxy to inject frame ordering, timeout, backpressure, and close behavior;
• assert frame-level events, terminal states, trace propagation, and drop reasons.

This makes semantic compatibility observable before a higher-level SDK or adapter masks it.

Non-goals

• Do not optimize vLLM chat completion by adding another wrapper around the same HTTP/SSE path.
• Do not promise cache wins for dynamic scenes or prompts without measured reuse.
• Do not split coarse Rust FFI work into many tiny boundary calls.
• Do not make transport packages or profile packages act as empty switches over hidden behavior.

Exit Criteria

• Runtime control and object profile baselines exist in the protocol documentation.
• Wire-level conformance schemas, target manifests, execution plans, and result reports exist.
• The low-code conformance generator can emit wire target manifests.
• At least one SDK/runtime can be tested by the wire runner without using its own adapter as the semantic oracle.