generated: '2026-08-26'
method: searched
source: https://docs.mthreads.com/asr/recording-recognition/introduction
name: Moore Threads API Conventions
description: >-
  Cross-cutting runtime semantics for the Moore Threads public API surface, read from the AIBook
  speech API references and the KUAE Cloud Coding Plan docs. The AIBook surfaces share one house
  style — a JSON envelope with task_id / status / status_text, binary audio on the wire, and an
  explicit Start/Send/Stop state machine — while KUAE Cloud inherits the Anthropic and OpenAI
  conventions wholesale because it implements those two protocols.
auth_style:
  summary: Opaque token. Header on HTTP surfaces, query parameter on WebSocket surfaces, request-body field on streaming TTS.
  see: authentication/moore-threads-authentication.yml
idempotency:
  supported: false
  header: null
  note: >-
    No idempotency key, no request-deduplication guarantee and no safe-retry contract is documented
    on any surface. The recording-file ASR flow is the one place where an accidental replay is
    visible to the caller: sending UploadDone twice returns 4015 and sending requests out of
    sequence returns 4016, which are guards against duplicate submission rather than idempotency.
pagination:
  supported: false
  note: No collection endpoint on any published surface, so nothing paginates.
field_expansion:
  supported: false
  note: >-
    The closest analogue is opt-in response enrichment on ASR — show_words, show_confidence,
    enable_speaker_info, enable_timestamp_alignment, show_intermediate_result — and on TTS —
    enable_subtitle, enable_phoneme_timestamp. Docs warn that the TTS timestamp flags raise
    first-packet latency on large-model voices.
metadata:
  supported: false
request_id_tracing:
  supported: partial
  field: task_id
  note: >-
    Every AIBook job/session gets a server-generated task_id echoed on every subsequent frame; it
    is the only correlation handle. There is no X-Request-Id header. The KUAE Cloud authentication
    error carries a "Request id:0217xxxxxxx" inside the message string, so a request id exists on
    that surface but is not exposed as a field or header.
versioning:
  style: uri-path
  values:
  - /api/v1/tts/generate
  - /api/v2/tts/stream_generate
  - /api/v1/asr
  - /api/v1/streaming_vc
  note: >-
    Path versioning, and the TTS pair demonstrates it is real — v1 is the unary HTTP surface and
    v2 is the streaming WebSocket surface. No version header, no dated version pin, no published
    policy for how long a version lives. KUAE Cloud has no Moore Threads version namespace of its
    own; /v1 there is the OpenAI protocol's own path.
error_envelope:
  style: vendor-numeric-in-body
  see: errors/moore-threads-error-codes.yml
rate_limit_signaling:
  headers: []
  note: >-
    No RateLimit-*, X-RateLimit-* or Retry-After header is documented on any surface. Exhaustion is
    signalled in-band by status 2003 (too many requests) and 2004 (service overload). The KUAE
    Cloud Coding Plan quota is a 5-hour rolling prompt budget with no runtime signal at all — the
    docs tell the human to wait for the next 5-hour window.
  see: rate-limits/moore-threads-rate-limits.yml
streaming:
  protocols:
  - websocket
  - http-chunked
  - grpc
  note: >-
    WebSocket is the primary streaming transport (TTS v2, realtime ASR, voice conversion). The
    on-premise LiteTTS package instead exposes gRPC StreamingSynthesizeSpeech on port 51061. KUAE
    Cloud streams via the Anthropic/OpenAI SSE conventions.
  state_machines:
  - api: Speech Recognition (realtime)
    sequence: StartTranscription -> binary audio frames -> StopTranscription -> TranscriptionCompleted
    constraint: Audio frames must arrive at least every 10 seconds or the server closes the session (4009).
  - api: Streaming Voice Conversion
    sequence: StartConversion -> binary audio frames -> StopConversion -> ConversionCompleted
    constraint: Send exactly 1000 ms of audio every 1000 ms; faster than realtime triggers 2004.
  - api: Recording-file Speech Recognition
    sequence: submit -> (upload chunks -> upload done) -> query | callback
    constraint: Chunks should not exceed 10 MB; each chunk must be acknowledged before the next is sent.
reversibility:
  grade: na
  applicable: false
  rationale: >-
    Nothing on the published Moore Threads API surface creates durable, externally-visible state
    that a caller could need to take back. TTS and voice conversion are pure transformations; ASR
    creates a transcription job whose only artifact is a result the server itself discards after
    72 hours; the KUAE Cloud Coding Plan is a chat-completions endpoint. There is no create/update/
    delete of a customer record, no payment, no provisioning call.
  write_surfaces:
  - operation: ASR submit (recording-file recognition)
    creates: A transcription task identified by task_id
    reversal_operation: null
    window: null
    note: >-
      No cancel, abort or delete operation is documented for a submitted task. The docs describe
      only forward motion — submit, upload, upload done, query. A task cannot be recalled once
      UploadDone is sent; the caller's only remedy is to stop polling and let the 72-hour retention
      expire the result. This is an honest gap, not an inferred window.
  dry_run_mode: na
  idempotency: na
  checked: '2026-08-26'
cross_links:
  errors: errors/moore-threads-error-codes.yml
  lifecycle: lifecycle/moore-threads-lifecycle.yml
  authentication: authentication/moore-threads-authentication.yml
  rate_limits: rate-limits/moore-threads-rate-limits.yml