generated: '2026-09-01'
method: searched
source:
- https://raw.githubusercontent.com/aws/api-models-aws/main/models/bedrock-runtime/service/2023-09-30/bedrock-runtime-2023-09-30.json
- https://docs.aws.amazon.com/nova/latest/userguide/using-invoke-api.html
- https://docs.aws.amazon.com/nova/latest/userguide/using-converse-api.html
- https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_StartAsyncInvoke.html
protocol:
  name: aws.protocols#restJson1
  transports: [http/1.1, h2]
  event_stream: vnd.amazon.eventstream over h2 and http/1.1
authentication:
  style: aws-sigv4
  signing_name: bedrock
  alternative: >-
    The service model declares smithy.api#httpBearerAuth alongside aws.auth#sigv4, which is the
    Amazon Bedrock API key (bearer token) path. See authentication/amazon-nova-authentication.yml.
  cross_link: authentication/amazon-nova-authentication.yml
idempotency:
  supported: partial
  mechanism: request-body token
  field: clientRequestToken
  applies_to: [startAsyncInvoke]
  smithy_trait: smithy.api#idempotencyToken
  header: null
  scope: per-account, per-operation
  retention: not published
  conflict_error: ConflictException (HTTP 400)
  note: >-
    StartAsyncInvoke is the one Amazon Nova write operation with a real idempotency contract:
    `clientRequestToken` is marked @idempotencyToken in Amazon's own Smithy model, and AWS SDKs
    auto-generate it when the caller omits it. The synchronous inference operations —
    InvokeModel, InvokeModelWithResponseStream, Converse, ConverseStream — carry NO idempotency
    key at all. A retried Converse call is a second billed generation, and an agent must treat it
    that way.
pagination:
  style: cursor
  request_params: [maxResults, nextToken]
  response_fields: [nextToken]
  applies_to: [ListAsyncInvokes]
  note: Only the async-invoke listing paginates; inference operations return a single response.
field_expansion:
  supported: false
metadata:
  supported: true
  mechanism: >-
    Converse accepts requestMetadata (a key/value map) for correlating an invocation with caller-side
    context; model invocation logging can deliver it to Amazon S3 / CloudWatch Logs.
request_tracing:
  request_id_header: x-amzn-RequestId
  error_type_header: x-amzn-ErrorType
  note: >-
    Every Bedrock Runtime response carries x-amzn-RequestId; that value is what AWS Support and
    CloudTrail correlate on.
versioning:
  style: service-version + model-id
  cross_link: lifecycle/amazon-nova-lifecycle.yml
error_envelope:
  format: aws-json-error
  body_fields: [message]
  type_header: x-amzn-ErrorType
  rfc9457: false
  cross_link: errors/amazon-nova-problem-types.yml
rate_limit_signaling:
  response_headers: []
  status_on_exhaustion: 429
  error_types: [ThrottlingException, ModelNotReadyException]
  note: >-
    Amazon Bedrock Runtime publishes NO RateLimit-* or X-RateLimit-* response headers and no
    Retry-After. An agent gets a 429 with x-amzn-ErrorType and nothing else — remaining quota,
    reset window and limit are only visible out-of-band through Service Quotas and CloudWatch.
    This is the weakest part of Nova's runtime contract for an autonomous caller.
  cross_link: rate-limits/amazon-nova-rate-limits.yml
dry_run_mode:
  supported: partial
  mechanism: >-
    CountTokens (POST /model/{modelId}/count-tokens) is marked @readonly in the AWS service model
    and lets a caller price and size a request before spending an inference call on it. There is
    no general dry-run flag on the inference operations themselves.
reversibility:
  grade: na
  applies: false
  rationale: >-
    Amazon Nova's write surface produces model output, not durable provider-side state that a
    caller can take back. There is no cancel, refund, void, reverse or undo operation anywhere in
    the Bedrock Runtime contract: the eleven operations in Amazon's own Smithy model are
    ApplyGuardrail, Converse, ConverseStream, CountTokens, GetAsyncInvoke, InvokeGuardrailChecks,
    InvokeModel, InvokeModelWithBidirectionalStream, InvokeModelWithResponseStream,
    ListAsyncInvokes and StartAsyncInvoke — and none of them reverses another.
  write_surfaces:
  - operation: invokeModel
    reversal: null
    window: null
    note: A completed generation is billed and cannot be un-billed. Nothing persists server-side to delete.
  - operation: converse
    reversal: null
    window: null
  - operation: converseStream
    reversal: null
    window: null
  - operation: invokeModelWithResponseStream
    reversal: null
    window: null
  - operation: startAsyncInvoke
    reversal: null
    window: null
    note: >-
      The closest thing to reversal in the whole surface, and it still is not one. A long-running
      Nova Reel job started with StartAsyncInvoke has NO stop or cancel operation in the Bedrock
      Runtime API — GetAsyncInvoke and ListAsyncInvokes only observe it. The only genuine
      protection is the clientRequestToken idempotency key, which prevents a duplicate job rather
      than undoing one. Output objects written to the caller's own Amazon S3 bucket are deletable
      by the caller, but that is S3, not Nova.
  docs: https://docs.aws.amazon.com/bedrock/latest/APIReference/API_Operations_Amazon_Bedrock_Runtime.html
  note: >-
    No window is asserted here because the provider states none. Recorded as `na` rather than a
    zero: an agent should plan for Nova calls being irreversible by design, not assume an undo it
    will not find.