generated: '2026-07-17'
method: searched
source: https://docs.rlmesh.dev/en/v0.1.0-rc.2/  (2026.06 workflow edition contract)
description: >-
  Cross-cutting request/response semantics for the RLMesh gRPC wire protocol,
  captured from the sealed 2026.06 workflow-edition behavioral contract. RLMesh
  is a bidirectional-streaming gRPC protocol (not HTTP/REST), so conventions are
  expressed in terms of Join-stream requests rather than HTTP headers.
authentication:
  style: none-at-protocol-level
  note: >-
    The 2026.06 edition explicitly does not constrain transport security or
    authentication (out of scope). Deployments layer their own transport
    security. The SAI platform CLI (separate surface) authenticates with a
    bearer API token — see authentication/ai-arena-authentication.yml.
request_tracing:
  field: request_id
  note: Every request on a Join stream is answered exactly once, and each response mirrors the originating request_id, so clients demux responses by id.
ordering:
  guarantee: strict-arrival-order-per-stream
  note: >-
    Env-service requests are processed strictly in arrival order, one at a time.
    A pipelining model service may complete requests for different env_ids in any
    order (advertised via the rlmesh.model.concurrent_predict.v1 capability), but
    per-env_id ordering is always preserved.
idempotency:
  supported: partial
  mechanism: operation-level (not an idempotency-key retry contract)
  note: >-
    ResolveAdapter is an idempotent upsert that fixes an env's EnvContract.
    Episode and env ids are UUIDv7 (never repeat), so a missed or late
    ResetAdapter only leaks memory and can never alias a new episode. There is no
    documented idempotency-key header for safe retry of arbitrary calls.
versioning:
  style: negotiated-workflow-edition + frozen-protocol-generation
  protocol_generation: rlmesh-wire-v1
  workflow_edition: '2026.06'
  handshake: >-
    Clients declare their protocol_generation and supported_workflow_editions at
    Handshake; generation compatibility is the only handshake-level decision, and
    the runtime reconciles the session edition as the floor across env, model, and
    runtime.
error_envelope:
  shape: in-band gRPC error messages (EnvError / ModelError)
  fields: [code, message, is_recoverable]
  semantics: >-
    A recoverable error leaves the session usable for further requests; a
    non-recoverable error means the client must abandon the session. Rejected
    actions are recoverable (InvalidAction); rejected observations are
    non-recoverable (Internal).
timeouts:
  request_field: timeout_ms
  client_options: [connect_timeout_seconds, request_timeout_seconds]
  note: A positive timeout_ms is a server-enforced deadline; expiry produces an in-band error with code TIMEOUT.
vectorization:
  note: >-
    A served environment is a fixed-width vector of num_envs sub-environments;
    observations/actions are batched SpaceValues; rewards, terminated_mask, and
    truncated_mask carry one entry per sub-environment in index order.
value_conformance:
  policy_env: RLMESH_VALIDATION_POLICY   # warn (default) | strict | off
  note: >-
    Out-of-bounds Box/Text values warn (reported under the rlmesh.conformance.warning
    info key) rather than fail by default; structural problems (wrong shape, dtype,
    arity, domain, missing key) and NaN are always rejected.
cross_links:
  errors: null   # no HTTP error catalog; errors are in-band gRPC EnvError/ModelError
  lifecycle: lifecycle/ai-arena-lifecycle.yml
  authentication: authentication/ai-arena-authentication.yml
  grpc: grpc/ai-arena-grpc.yml