dd-trace 6.16.0 → 6.18.0

package/packages/datadog-instrumentations/src/openai-realtime/session.js

@@ -0,0 +1,998 @@
+'use strict'
+
+const { randomUUID } = require('node:crypto')
+
+const {
+  bytesPerSecond,
+  realtimeAudioFormatToMime,
+  segmentDurationMs,
+} = require('../../../dd-trace/src/llmobs/audio-codec')
+const log = require('../../../dd-trace/src/log')
+const { extractResponseTools, normalizeResponseEventType } = require('./events')
+const { InputTurn, ResponseTurn } = require('./turns')
+
+// Realtime PCM is 24 kHz mono by spec; overridden from the session's format object when present.
+const DEFAULT_AUDIO_RATE = 24_000
+
+// Ceiling on how long a finished turn is held waiting for its audio to finish playing. A connection
+// that goes idle right after a response is the gap: nothing fires, so the held turn would wait for
+// the next event or for close. A timed flush is deliberately avoided — it would finalize turns off
+// the caller's thread, and this state machine is only safe because every path runs on it.
+const PARK_MAX_MS = 5000
+
+/**
+ * @typedef {import('./turns').ResponseTurn} Turn
+ * @typedef {import('./turns').ToolCall} ToolCall
+ * @typedef {import('./turns').ToolResult} ToolResult
+ *
+ * @typedef {{
+ *   startTime: number,
+ *   finishTime: number,
+ *   transcript: string,
+ * }} SpeechWindow
+ *
+ * @typedef {{
+ *   text: string,
+ *   transcript: string,
+ *   audio: Buffer,
+ *   audioPresent: boolean,
+ *   mimeType: string,
+ *   sampleRate: number,
+ *   toolCalls?: ToolCall[],
+ *   toolResults: ToolResult[],
+ * }} TurnSide
+ *
+ * @typedef {{
+ *   sessionId: string,
+ *   model: string | undefined,
+ *   basePath: string,
+ *   metadata: Record<string, unknown>,
+ *   usage: object | undefined,
+ *   failed: boolean,
+ *   error: { type?: string, code?: string, message?: string } | undefined,
+ *   runInContext: ((fn: () => void) => void) | undefined,
+ *   root: { startTime: number, finishTime: number },
+ *   llm: { startTime: number, finishTime: number },
+ *   userSpeech: SpeechWindow | undefined,
+ *   agentSpeech: SpeechWindow | undefined,
+ *   input: TurnSide,
+ *   output: TurnSide,
+ * }} TurnDescriptor
+ */
+
+/**
+ * Coerce a value the server sent to a finite number, or `undefined`.
+ *
+ * @param {unknown} value
+ * @returns {number | undefined}
+ */
+function toFiniteNumber (value) {
+  const number = Number(value)
+  return Number.isFinite(number) ? number : undefined
+}
+
+/**
+ * Flatten the provider's failure detail to plain strings, or `undefined` when it said nothing
+ * usable. Kept to the three fields OpenAI documents so a span carries the type and message a
+ * responder actually needs, without copying an arbitrary provider object onto a tag.
+ *
+ * @param {unknown} error
+ */
+function providerError (error) {
+  if (error === null || typeof error !== 'object') return
+
+  /** @type {{ type?: string, code?: string, message?: string }} */
+  const flattened = {}
+  let reported = false
+  for (const field of ['type', 'code', 'message']) {
+    if (error[field] == null) continue
+    flattened[field] = String(error[field])
+    reported = true
+  }
+
+  return reported ? flattened : undefined
+}
+
+/**
+ * The audio format to interpret a segment with: the one recorded when its first frame arrived,
+ * falling back to the session's current format.
+ *
+ * The fallback covers a segment that never recorded one — it holds no audio, or the session had not
+ * announced a format yet — and is not the mutable-format case: a segment that did record a format
+ * keeps it, so a later `session.update` cannot retime bytes that arrived under the old one.
+ *
+ * @param {import('./audio-accumulator')} audio
+ * @param {string} mimeType
+ * @param {number} sampleRate
+ * @returns {{ mimeType: string, sampleRate: number }}
+ */
+function segmentFormat (audio, mimeType, sampleRate) {
+  return { mimeType: audio.mimeType || mimeType, sampleRate: audio.sampleRate || sampleRate }
+}
+
+/**
+ * Drives per-turn spans off one realtime connection's event stream.
+ *
+ * Turns are accumulated as plain data and replayed as a back-dated span tree at finalize, so nothing
+ * is held open across a deferred finalize (a late input transcription, or audio still playing) and a
+ * dropped connection cannot leak an unfinished span.
+ */
+class RealtimeSession {
+  /** Per-connection id grouping every turn of this conversation in the UI. */
+  #sessionId = randomUUID().replaceAll('-', '')
+
+  /** @type {(descriptor: TurnDescriptor) => void} */
+  #emitTurn
+
+  /** @type {(turn: Turn) => void} */
+  #captureContext
+
+  /** @type {string | undefined} */
+  #model
+
+  /** @type {string} */
+  #basePath
+
+  /** @type {Record<string, unknown>} */
+  #sessionConfig = {}
+
+  #inputAudioMime = ''
+  #outputAudioMime = ''
+  #inputAudioRate = DEFAULT_AUDIO_RATE
+  #outputAudioRate = DEFAULT_AUDIO_RATE
+
+  /**
+   * Whether the session enabled input-audio transcription. A turn's finalize is only deferred to
+   * wait for a transcript when one is actually configured — otherwise none is ever coming.
+   */
+  #inputTranscriptionEnabled = false
+
+  /**
+   * Offset (ms) of the end of all input audio appended this session, and the wall clock (epoch ms)
+   * at which we reached it. Together they place a VAD event's buffer offset on the wall clock.
+   */
+  #inputBufferMs = 0
+  /** @type {number | undefined} */
+  #inputBufferMsAt = undefined
+
+  /**
+   * Bytes appended before the audio format was known — a client streaming from its own thread can
+   * beat `session.created`. Held as a backlog and folded into the clock once a rate arrives, rather
+   * than writing the origin off and disabling onset projection for the whole session.
+   */
+  #pendingInputBytes = 0
+
+  /**
+   * Whether anything consumes the captured audio, asked once per segment rather than held as a
+   * snapshot. Only LLM Observability does, so with it disabled — the default — the segments keep
+   * their byte counts, which size the speech windows, and hold no bytes. See `AudioAccumulator`.
+   *
+   * A predicate because the answer can change while a connection is open: `tracer.use('openai', {
+   * llmobs: false })` unsubscribes mid-session, and a snapshot taken at connect would keep buffering
+   * megabytes for a consumer that had gone away (or keep a session that gained one unable to
+   * capture). `AudioAccumulator` resolves it on the frame that opens a segment, which is as late as
+   * possible while still being once per segment: asking per frame could flip retention mid-clip and
+   * leave a segment holding only part of itself, while resolving at construction would be far too
+   * early — a pending input segment is built as soon as the previous turn starts.
+   *
+   * @type {() => boolean}
+   */
+  #shouldRetainAudio
+
+  /** @type {InputTurn} */
+  #pendingInput
+
+  /** @type {Map<string, Turn>} */
+  #responses = new Map()
+
+  /** @type {Map<string, string>} */
+  #inputTranscripts = new Map()
+
+  /** call_id -> function name, so a later `function_call_output` can be labeled with its tool name. */
+  #toolCallNames = new Map()
+
+  /**
+   * Turns whose response is done but whose input transcription hasn't arrived yet.
+   *
+   * @type {Turn[]}
+   */
+  #awaiting = []
+
+  /**
+   * Finished turns held open while their audio is still playing, so a barge-in truncation can still
+   * cap them.
+   *
+   * @type {Turn[]}
+   */
+  #playing = []
+
+  /**
+   * Whether this connection's client has ever truncated — which is what makes holding turns open
+   * worth its cost. A client either implements barge-in or it does not.
+   */
+  #clientTruncates = false
+
+  #closed = false
+
+  /**
+   * @param {object} options
+   * @param {(descriptor: TurnDescriptor) => void} options.emitTurn
+   * @param {(turn: Turn) => void} options.captureContext
+   * @param {string} [options.model]
+   * @param {string} [options.basePath]
+   * @param {() => boolean} [options.shouldRetainAudio]
+   */
+  constructor ({ emitTurn, captureContext, model, basePath = '', shouldRetainAudio = () => true }) {
+    this.#emitTurn = emitTurn
+    this.#captureContext = captureContext
+    this.#model = model
+    this.#basePath = basePath
+    this.#shouldRetainAudio = shouldRetainAudio
+    this.#pendingInput = new InputTurn(shouldRetainAudio)
+  }
+
+  // -- event entry points ---------------------------------------------------
+
+  /**
+   * @param {Record<string, unknown>} event
+   * @param {number} now - Epoch ms at which the event was observed.
+   */
+  onClientEvent (event, now) {
+    try {
+      this.#flushPlaying(now)
+
+      switch (event?.type) {
+        // First: a server-VAD client appends microphone audio many times a second.
+        case 'input_audio_buffer.append':
+          if (event.audio) this.#appendInputAudio(event.audio, now)
+          break
+        case 'conversation.item.truncate':
+          // Client -> server: "the listener only got this far into that item."
+          this.#onTruncate(event.item_id, event.audio_end_ms, now)
+          break
+        case 'session.update':
+          this.#updateSessionConfig(event.session)
+          break
+        case 'input_audio_buffer.clear':
+          // Discarded input audio must not be attributed to the next response.
+          this.#pendingInput.discardAudio()
+          break
+        case 'conversation.item.create':
+          this.#absorbInputItem(event.item, now)
+          break
+      }
+    } catch (error) {
+      log.debug('Error handling OpenAI realtime client event: %s', error?.message)
+    }
+  }
+
+  /**
+   * @param {Record<string, unknown>} event
+   * @param {number} now - Epoch ms at which the event was observed.
+   */
+  onServerEvent (event, now) {
+    try {
+      this.#flushPlaying(now)
+
+      const eventType = event?.type
+      if (typeof eventType !== 'string') return
+
+      switch (eventType) {
+        case 'session.created':
+        case 'session.updated':
+          this.#updateSessionConfig(event.session)
+          return
+        case 'conversation.item.truncated':
+          // The server's acknowledgement of a client truncation. `capTo` is absolute, so handling
+          // both it and the client event applies the cap once.
+          this.#onTruncate(event.item_id, event.audio_end_ms, now)
+          return
+        case 'input_audio_buffer.speech_started':
+          this.#onSpeechStarted(event.audio_start_ms, now)
+          return
+        case 'input_audio_buffer.speech_stopped':
+          // The commit that follows is the authoritative end of user speech and overwrites this;
+          // recording it here only covers a session that never commits, so the window still gets an
+          // end.
+          this.#pendingInput.speechEndTime ??= now
+          return
+        case 'input_audio_buffer.committed':
+          this.#pendingInput.itemId = event.item_id == null ? undefined : String(event.item_id)
+          this.#pendingInput.speechEndTime = now
+          return
+        case 'input_audio_buffer.cleared':
+          this.#pendingInput.discardAudio()
+          return
+        case 'conversation.item.input_audio_transcription.completed':
+          this.#onInputTranscript(event.item_id, event.transcript, now)
+          return
+        case 'conversation.item.input_audio_transcription.failed':
+          // No transcription is coming for this item. Recorded as an empty transcript rather than
+          // just flushing whatever is already waiting: this can land *before* the response's
+          // `response.done`, and without the record `#finishResponse` would then park the turn for
+          // an event that has already happened.
+          this.#onInputTranscript(event.item_id, '', now)
+          return
+        case 'response.created':
+          this.#startResponse(event.response?.id ?? event.response_id, event.response, now)
+          return
+        case 'response.done':
+          this.#finishResponse(event.response?.id ?? event.response_id, event.response, now)
+          return
+        default:
+          this.#handleResponseDelta(event, eventType, now)
+      }
+    } catch (error) {
+      log.debug('Error handling OpenAI realtime server event: %s', error?.message)
+    }
+  }
+
+  /**
+   * Finalize everything still open. Idempotent, so every close path can call it.
+   *
+   * @param {number} now - Epoch ms.
+   * @param {boolean} [failed] - The connection ended abnormally. Only responses still in flight are
+   *   marked: a turn already awaiting a transcript or playback had its `response.done`, so it
+   *   succeeded whatever the transport did afterwards.
+   */
+  finishSession (now, failed = false) {
+    if (this.#closed) return
+    this.#closed = true
+
+    try {
+      this.#flushAwaiting(now)
+      this.#flushPlaying(now, true)
+
+      // In-flight turns that never saw `response.done` (closed mid-turn). Whatever partial data we
+      // have is submitted.
+      for (const turn of this.#responses.values()) {
+        if (failed) turn.status = 'failed'
+        this.#applyCachedTranscript(turn)
+        this.#finalizeTurn(turn, now, true)
+      }
+
+      this.#responses.clear()
+      this.#inputTranscripts.clear()
+      this.#toolCallNames.clear()
+      // Audio buffered for a turn that will now never start. An app that holds on to closed
+      // transport objects keeps their sessions reachable through the connection map, and a
+      // server-VAD client streams the microphone continuously, so this is up to the whole retention
+      // cap per closed connection with nothing left to consume it.
+      this.#pendingInput = new InputTurn(this.#shouldRetainAudio)
+    } catch (error) {
+      log.debug('Error finalizing OpenAI realtime session: %s', error?.message)
+    }
+  }
+
+  // -- response deltas ------------------------------------------------------
+
+  /**
+   * @param {Record<string, unknown>} event
+   * @param {string} eventType
+   * @param {number} now
+   */
+  #handleResponseDelta (event, eventType, now) {
+    const turn = this.#responses.get(String(event.response_id))
+    if (turn === undefined) return
+
+    switch (normalizeResponseEventType(eventType)) {
+      case 'response.audio.delta': {
+        const { delta, item_id: itemId } = event
+        if (!delta) return
+        if (itemId != null) {
+          // Remember where this item's audio starts in the turn's segment, before appending, so a
+          // truncation reported against the item maps onto the segment.
+          const key = String(itemId)
+          if (!turn.audioItemStarts.has(key)) turn.audioItemStarts.set(key, turn.audio.totalDecodedBytes)
+        }
+        turn.audio.append(delta, now, this.#outputAudioMime, this.#outputAudioRate)
+        return
+      }
+      case 'response.audio_transcript.delta':
+        turn.transcript.appendDelta(event.item_id, event.delta ?? '')
+        return
+      case 'response.audio_transcript.done':
+        if (event.transcript != null) turn.transcript.complete(event.item_id, String(event.transcript))
+        return
+      case 'response.text.delta':
+        turn.text.appendDelta(event.item_id, event.delta ?? '')
+        return
+      case 'response.text.done':
+        if (event.text != null) turn.text.complete(event.item_id, String(event.text))
+    }
+  }
+
+  // -- input audio and the speech window ------------------------------------
+
+  /**
+   * Buffer a client audio append for the pending turn and advance the input-buffer clock.
+   *
+   * @param {string} base64
+   * @param {number} now
+   */
+  #appendInputAudio (base64, now) {
+    const pending = this.#pendingInput
+
+    // Fold in anything buffered before the format was known first, so the base offset captured below
+    // sits on the same timeline the VAD offsets are later projected against.
+    this.#advanceInputBufferClock(0, now)
+
+    if (pending.audio.startTime === undefined && this.#inputBufferMsAt !== undefined) {
+      // First frame of this turn and the clock is live: remember where it sits on the session's
+      // input-buffer timeline, so a VAD offset can be turned into a byte offset into what we buffer
+      // here. Left unset while the clock is dead, since a base of 0 would read as "this turn starts
+      // at the session origin" and over-trim the front of the segment.
+      pending.audioBaseMs = this.#inputBufferMs
+    }
+
+    const decodedBytes = pending.audio.append(base64, now, this.#inputAudioMime, this.#inputAudioRate)
+    this.#advanceInputBufferClock(decodedBytes, now)
+  }
+
+  /**
+   * Track how far into the session's input audio the buffer now extends, and when we got there.
+   *
+   * @param {number} decodedBytes
+   * @param {number} now
+   */
+  #advanceInputBufferClock (decodedBytes, now) {
+    this.#pendingInputBytes += decodedBytes
+
+    const rate = bytesPerSecond(this.#inputAudioMime, this.#inputAudioRate)
+    // A format we can never rate leaves the projection unavailable: the backlog simply never
+    // converts and `#inputBufferMsAt` stays undefined, which is what marks the clock dead.
+    if (!rate) return
+
+    this.#inputBufferMs += this.#pendingInputBytes / rate * 1000
+    this.#pendingInputBytes = 0
+    this.#inputBufferMsAt = now
+  }
+
+  /**
+   * Anchor the pending turn's user-speech window on the VAD speech onset.
+   *
+   * A server-VAD client streams the microphone continuously, so the first buffer append of a turn
+   * lands the instant the *previous* turn was committed: it marks when we started listening, not
+   * when the human started speaking. Left at that, every user-speech window swallows the whole
+   * preceding agent response and consecutive turns overlap on the session timeline.
+   * `input_audio_buffer.speech_started` is the real onset, and the audio it points at
+   * (`audio_start_ms`, which already includes the session's `prefix_padding_ms`) is the audio worth
+   * keeping, so the buffered lead-in is trimmed off the front to keep the captured audio and the
+   * reported window in step.
+   *
+   * Only the first onset of a turn counts: a turn that VAD splits into several speech runs before a
+   * single commit is still one committed item, which began at the first run.
+   *
+   * @param {unknown} audioStartMs
+   * @param {number} now
+   */
+  #onSpeechStarted (audioStartMs, now) {
+    const pending = this.#pendingInput
+    if (pending.speechStartTime !== undefined) return
+
+    const onset = this.#bufferOffsetToWallTime(audioStartMs, now)
+    pending.speechStartTime = onset
+
+    pending.audio.trimLeading(this.#preOnsetBytes(audioStartMs))
+    // Re-anchor the segment on the onset: whatever survived the trim starts there. A trim that
+    // covered the whole segment resets it instead, and re-anchoring then would leave `startTime`
+    // set with no format recorded — `append` only records the format on the frame that opens a
+    // segment, so every later frame would skip it and the segment would fall back to whatever
+    // format the session holds at describe time.
+    if (pending.audio.startTime !== undefined) pending.audio.startTime = onset
+  }
+
+  /**
+   * Byte count of the audio buffered for this turn ahead of the speech onset.
+   *
+   * @param {unknown} audioStartMs
+   */
+  #preOnsetBytes (audioStartMs) {
+    const baseMs = this.#pendingInput.audioBaseMs
+    const onsetMs = toFiniteNumber(audioStartMs)
+    const rate = bytesPerSecond(this.#inputAudioMime, this.#inputAudioRate)
+    if (baseMs === undefined || onsetMs === undefined || !rate) return 0
+
+    const bytes = Math.max(0, Math.trunc((onsetMs - baseMs) / 1000 * rate))
+    // Cut on a sample boundary, as the truncation cap does: `audioBaseMs` is a fractional millisecond
+    // offset, so this lands on an odd byte often enough to matter, and slicing a PCM16 segment
+    // mid-sample pairs every low byte with the next sample's high byte — the whole clip decodes to
+    // noise. Rounding down keeps a byte of lead-in rather than corrupting what follows.
+    return bytes - bytes % 2 // keep PCM16 samples whole; a byte is nothing for G.711
+  }
+
+  /**
+   * Project an input-buffer offset — a VAD event's `audio_start_ms`, measured from the start of all
+   * audio written to the buffer this session — onto the wall clock.
+   *
+   * This assumes the client appends audio roughly in real time, which is true for a live microphone,
+   * the only case where a wall-clock speech window means anything. Knowing how much audio had been
+   * appended at a known instant, the offset is that instant minus the audio still ahead of it. The
+   * result is clamped to a window we can defend — no earlier than this turn's first buffered frame,
+   * no later than when we observed the event — so a bursty or pre-recorded sender degrades to a sane
+   * bound instead of a wild timestamp, and falls back to the observation time when the projection is
+   * unavailable.
+   *
+   * @param {unknown} offsetMs
+   * @param {number} observedTime
+   */
+  #bufferOffsetToWallTime (offsetMs, observedTime) {
+    const offset = toFiniteNumber(offsetMs)
+    if (offset === undefined || this.#inputBufferMsAt === undefined) return observedTime
+
+    let projected = this.#inputBufferMsAt - (this.#inputBufferMs - offset)
+
+    const earliest = this.#pendingInput.audio.startTime
+    if (earliest !== undefined) projected = Math.max(projected, earliest)
+
+    return Math.min(projected, observedTime)
+  }
+
+  // -- barge-in (agent playback cut short) ----------------------------------
+
+  /**
+   * Cap an assistant audio segment at what the listener actually heard.
+   *
+   * Over a WebSocket the client owns playback and the model streams audio faster than it plays, so
+   * on a barge-in the client stops its speaker and reports how far it got. Audio delivered past that
+   * point was never heard; without this the stored agent audio — and the agent-speech window derived
+   * from it — covers the whole generated response and runs past the interruption into the next user
+   * turn.
+   *
+   * Seeing a truncation also marks this connection's client as one that cuts playback short, which
+   * is what makes holding its turns open worthwhile.
+   *
+   * @param {unknown} itemId
+   * @param {unknown} audioEndMs
+   * @param {number} now
+   */
+  #onTruncate (itemId, audioEndMs, now) {
+    this.#clientTruncates = true
+
+    const endMs = toFiniteNumber(audioEndMs)
+    if (itemId == null || endMs === undefined) return
+
+    const item = String(itemId)
+    for (const turn of this.#openTurns()) {
+      const itemStart = turn.audioItemStarts.get(item)
+      if (itemStart === undefined) continue
+
+      const { mimeType, sampleRate } = segmentFormat(turn.audio, this.#outputAudioMime, this.#outputAudioRate)
+      const rate = bytesPerSecond(mimeType, sampleRate)
+      if (!rate) return
+
+      const cap = itemStart + Math.trunc(endMs / 1000 * rate)
+      turn.audio.capTo(cap - cap % 2) // keep PCM16 samples whole; a byte is nothing for G.711
+
+      const playingIndex = this.#playing.indexOf(turn)
+      if (playingIndex !== -1) {
+        // Playback ended when the listener cut it off, so stop waiting on it.
+        this.#playing.splice(playingIndex, 1)
+        this.#finalizeTurn(turn, now, true)
+      }
+      return
+    }
+  }
+
+  /**
+   * Every turn we could still amend: in flight, awaiting a transcript, or awaiting playback.
+   *
+   * @returns {Turn[]}
+   */
+  #openTurns () {
+    return [...this.#responses.values(), ...this.#awaiting, ...this.#playing]
+  }
+
+  /**
+   * Hold a finished turn while its audio is still playing, so a late truncation can still cap it.
+   *
+   * `response.done` normally lands mid-playback (generation outruns playback), and a barge-in
+   * truncation arrives after that — too late for a turn we already submitted. Holding costs
+   * submission latency and a wider window in which to lose the turn if the process dies, so we only
+   * hold on connections whose client has actually truncated before. Clients that never truncate hear
+   * every byte we captured and finalize at `response.done` exactly as before, paying nothing. The
+   * cost of that trade is that the first interruption on a connection is reported untruncated.
+   *
+   * Reports whether the turn was parked, so the caller knows not to finalize it.
+   *
+   * @param {Turn} turn
+   * @param {number} now
+   */
+  #parkForPlayback (turn, now) {
+    if (!this.#clientTruncates || turn.audio.startTime === undefined) return false
+
+    const { mimeType, sampleRate } = segmentFormat(turn.audio, this.#outputAudioMime, this.#outputAudioRate)
+    const playbackMs = segmentDurationMs(turn.audio.totalDecodedBytes, mimeType, sampleRate)
+    if (playbackMs === undefined) return false
+
+    const endTime = turn.audio.startTime + playbackMs
+    if (endTime <= now) return false
+
+    turn.playbackEndTime = Math.min(endTime, now + PARK_MAX_MS)
+    this.#playing.push(turn)
+    return true
+  }
+
+  /**
+   * Finalize held turns whose audio has finished playing, or all of them when forced.
+   *
+   * Event-driven rather than timed: a realtime connection is chatty — a streaming client appends
+   * microphone audio continuously — so this runs often enough to submit a turn shortly after its
+   * playback ends, and the next turn and connection close both force it so a turn can't leak.
+   *
+   * @param {number} now
+   * @param {boolean} [force]
+   */
+  #flushPlaying (now, force = false) {
+    if (this.#playing.length === 0) return
+
+    for (let i = this.#playing.length - 1; i >= 0; i--) {
+      const turn = this.#playing[i]
+      if (!force && turn.playbackEndTime !== undefined && turn.playbackEndTime > now) continue
+
+      this.#playing.splice(i, 1)
+      this.#finalizeTurn(turn, now, true)
+    }
+  }
+
+  // -- turn lifecycle -------------------------------------------------------
+
+  /**
+   * @param {unknown} responseId
+   * @param {Record<string, unknown> | undefined} response - The created response, which says which
+   *   conversation it belongs to.
+   * @param {number} now
+   */
+  #startResponse (responseId, response, now) {
+    if (responseId == null) return
+
+    // The server tells us directly: a response created with `conversation: 'none'` reports a null
+    // `conversation_id`, while one in the default conversation — including every response server VAD
+    // creates on its own — reports an id. Only the latter owns the buffered user input.
+    //
+    // Read off `response.created` rather than correlated back to the client's `response.create`.
+    // Position-matching the two streams cannot survive a create the server rejects (no
+    // `response.created` ever arrives) or several in flight at once, and both of those
+    // mis-classifications cost a real turn its user speech.
+    //
+    // `conversation_id` is absent rather than null on the beta realtime surface, which is why this
+    // tests for null exactly: an unknown conversation is treated as the conversation, matching how
+    // that surface behaved before out-of-band responses were modelled at all.
+    const outOfBand = response?.conversation_id === null
+
+    // Classify before flushing: an out-of-band response runs *alongside* the conversation, so it is
+    // not evidence that a pending turn's transcript has stopped coming.
+
+    if (!outOfBand) {
+      // A new conversational turn starting means a prior turn's input transcription is almost
+      // certainly not coming anymore, and that any held playback is over (or was cut off without a
+      // truncation reaching us), so flush both rather than let a turn hang.
+      this.#flushAwaiting(now)
+      this.#flushPlaying(now, true)
+    }
+
+    const input = outOfBand ? new InputTurn(this.#shouldRetainAudio) : this.#pendingInput
+    const turn = new ResponseTurn(input, now, this.#shouldRetainAudio)
+    if (!outOfBand) this.#pendingInput = new InputTurn(this.#shouldRetainAudio)
+    turn.model = this.#model
+
+    this.#captureContext(turn)
+    this.#responses.set(String(responseId), turn)
+  }
+
+  /**
+   * @param {unknown} responseId
+   * @param {Record<string, unknown>} response
+   * @param {number} now
+   */
+  #finishResponse (responseId, response, now) {
+    if (responseId == null) return
+
+    const key = String(responseId)
+    const turn = this.#responses.get(key)
+    if (turn === undefined) return
+    this.#responses.delete(key)
+
+    turn.responseDoneTime = now
+    turn.usage = response?.usage
+    turn.model = response?.model || turn.model || this.#model
+    turn.status = response?.status
+    // The provider's actionable detail for a failure. Without it a failed turn reaches the backend
+    // as a bare `error: 1`, which says a realtime call broke but nothing about why.
+    turn.error = providerError(response?.status_details?.error)
+
+    const { toolCalls, toolResults } = extractResponseTools(response)
+    turn.toolCalls = toolCalls
+    turn.toolResults = toolResults
+    // Remember each function call's name so the `function_call_output` the app returns later can be
+    // labeled with it — the output event itself only carries the call_id.
+    for (const toolCall of toolCalls) {
+      if (toolCall.type === 'function' && toolCall.toolId) {
+        this.#toolCallNames.set(toolCall.toolId, toolCall.name)
+      }
+    }
+
+    this.#applyCachedTranscript(turn)
+
+    // Hold the turn open for a late input transcription ONLY when transcription is actually enabled
+    // and this item has not already reached a terminal state: otherwise no transcript is ever
+    // coming, and waiting would needlessly delay every turn until the next one, and the last turn
+    // until close. A cached entry — including the empty string a completion-with-no-text or a
+    // failure records — is that terminal state, and it is why the check is `has` rather than the
+    // transcript's truthiness.
+    if (!turn.input.transcript &&
+        turn.input.itemId !== undefined &&
+        this.#inputTranscriptionEnabled &&
+        !this.#inputTranscripts.has(turn.input.itemId)) {
+      this.#awaiting.push(turn)
+      return
+    }
+
+    this.#finalizeTurn(turn, now)
+  }
+
+  /**
+   * @param {unknown} itemId
+   * @param {unknown} transcript
+   * @param {number} now
+   */
+  #onInputTranscript (itemId, transcript, now) {
+    const item = itemId == null ? undefined : String(itemId)
+    const text = transcript == null ? '' : String(transcript)
+
+    if (item !== undefined) this.#inputTranscripts.set(item, text)
+    if (this.#pendingInput.itemId === item && !this.#pendingInput.transcript) {
+      this.#pendingInput.transcript = text
+    }
+
+    // A finished turn may have been waiting on exactly this transcript — finalize it now.
+    this.#finalizeAwaitingFor(itemId, now, text)
+  }
+
+  /**
+   * @param {unknown} itemId
+   * @param {number} now
+   * @param {string} [transcript]
+   */
+  #finalizeAwaitingFor (itemId, now, transcript) {
+    if (itemId == null) return
+    const item = String(itemId)
+
+    for (let i = this.#awaiting.length - 1; i >= 0; i--) {
+      const turn = this.#awaiting[i]
+      if (turn.input.itemId !== item) continue
+
+      if (transcript) turn.input.transcript ||= transcript
+      this.#awaiting.splice(i, 1)
+      this.#finalizeTurn(turn, now)
+    }
+  }
+
+  /**
+   * @param {number} now
+   */
+  #flushAwaiting (now) {
+    if (this.#awaiting.length === 0) return
+
+    const awaiting = this.#awaiting
+    this.#awaiting = []
+    for (const turn of awaiting) this.#finalizeTurn(turn, now, true)
+  }
+
+  /**
+   * @param {Turn} turn
+   */
+  #applyCachedTranscript (turn) {
+    if (turn.input.transcript || turn.input.itemId === undefined) return
+    turn.input.transcript = this.#inputTranscripts.get(turn.input.itemId) ?? ''
+  }
+
+  /**
+   * @param {Turn} turn
+   * @param {number} now
+   * @param {boolean} [force] - Skip parking, for the paths that must not wait (close, next turn).
+   */
+  #finalizeTurn (turn, now, force = false) {
+    // The turn's data is complete, but on a barge-in-capable client the agent's audio may still be
+    // playing and a truncation may yet cut it short — hold the turn rather than submit audio the
+    // listener might never hear.
+    if (!force && this.#parkForPlayback(turn, now)) return
+
+    // Drop the cached transcript for this turn's input item so the map can't grow across a long
+    // session. Every finalize path goes through here.
+    if (turn.input.itemId !== undefined) this.#inputTranscripts.delete(turn.input.itemId)
+
+    try {
+      this.#emitTurn(this.#describeTurn(turn, now))
+    } catch (error) {
+      log.debug('Error emitting OpenAI realtime turn spans: %s', error?.message)
+    }
+  }
+
+  /**
+   * Flatten a finished turn into the boundaries and payloads the plugins need, so they never reach
+   * into this class's state.
+   *
+   * @param {Turn} turn
+   * @param {number} now
+   * @returns {TurnDescriptor}
+   */
+  #describeTurn (turn, now) {
+    const { input } = turn
+
+    // The VAD speech onset when we have it. The first buffered frame only approximates the onset,
+    // for a client that appends audio solely while the user talks (client-side turn detection).
+    const userStart = input.speechStartTime ?? input.audio.startTime
+    const inputFormat = segmentFormat(input.audio, this.#inputAudioMime, this.#inputAudioRate)
+    const outputFormat = segmentFormat(turn.audio, this.#outputAudioMime, this.#outputAudioRate)
+    const inputDurationMs = segmentDurationMs(
+      input.audio.totalDecodedBytes, inputFormat.mimeType, inputFormat.sampleRate
+    )
+    const userEnd = input.speechEndTime ??
+      (userStart !== undefined && inputDurationMs !== undefined ? userStart + inputDurationMs : undefined)
+
+    const agentStart = turn.audio.startTime
+    const playbackMs = segmentDurationMs(
+      turn.audio.totalDecodedBytes, outputFormat.mimeType, outputFormat.sampleRate
+    )
+    const agentEnd = agentStart === undefined
+      ? undefined
+      : (playbackMs === undefined ? (turn.responseDoneTime ?? now) : agentStart + playbackMs)
+
+    // The llm span measures model work: it opens at the end of user speech, not when the human
+    // started talking, and closes when generation completes.
+    const llmStartTime = input.speechEndTime ?? input.audio.startTime ?? turn.createdTime
+    const llmFinishTime = Math.max(turn.responseDoneTime ?? now, llmStartTime)
+
+    const rootStartTime = userStart ?? input.speechEndTime ?? turn.createdTime
+    // Unlike dd-trace-py, clamp the root to its children's actual ends so it always contains them: a
+    // turn finalized without `response.done` would otherwise finish before its own llm span.
+    const rootFinishTime = Math.max(rootStartTime, llmFinishTime, agentEnd ?? 0, userEnd ?? 0)
+
+    return {
+      sessionId: this.#sessionId,
+      model: turn.model,
+      basePath: this.#basePath,
+      metadata: { ...this.#sessionConfig },
+      usage: turn.usage,
+      failed: turn.status === 'failed',
+      error: turn.error,
+      runInContext: turn.runInContext,
+      root: { startTime: rootStartTime, finishTime: rootFinishTime },
+      llm: { startTime: llmStartTime, finishTime: llmFinishTime },
+      userSpeech: userStart === undefined || userEnd === undefined
+        ? undefined
+        : {
+            startTime: userStart,
+            finishTime: Math.max(userStart, userEnd),
+            transcript: input.transcript || input.text,
+          },
+      agentSpeech: agentStart === undefined || agentEnd === undefined
+        ? undefined
+        : {
+            startTime: agentStart,
+            finishTime: Math.max(agentStart, agentEnd),
+            transcript: turn.transcript.value || turn.text.value,
+          },
+      input: {
+        text: input.text,
+        transcript: input.transcript,
+        audio: input.audio.toBuffer(),
+        audioPresent: input.audio.present,
+        mimeType: inputFormat.mimeType,
+        sampleRate: inputFormat.sampleRate,
+        toolResults: input.toolResults,
+      },
+      output: {
+        text: turn.text.value,
+        transcript: turn.transcript.value,
+        audio: turn.audio.toBuffer(),
+        audioPresent: turn.audio.present,
+        mimeType: outputFormat.mimeType,
+        sampleRate: outputFormat.sampleRate,
+        toolCalls: turn.toolCalls,
+        toolResults: turn.toolResults,
+      },
+    }
+  }
+
+  // -- config extraction ----------------------------------------------------
+
+  /**
+   * @param {Record<string, unknown>} session
+   */
+  #updateSessionConfig (session) {
+    if (session == null) return
+
+    if (typeof session.model === 'string' && session.model) this.#model = session.model
+    if (session.instructions != null) this.#sessionConfig.instructions = String(session.instructions)
+
+    const modalities = session.output_modalities ?? session.modalities
+    if (modalities) this.#sessionConfig.output_modalities = [...modalities]
+
+    const { audio } = session
+    let inputFormat = audio?.input?.format
+    let outputFormat = audio?.output?.format
+    let voice = audio?.output?.voice
+
+    // A `session.update` carrying the transcription field explicitly set to `null` turns
+    // transcription off, so track whether the field is *present* rather than whether it is set.
+    // Latching the flag on would leave every later turn deferred in `#awaiting` for a transcript
+    // that is never coming; ignoring absence is equally required, since a partial update that says
+    // nothing about transcription must not disable it.
+    if (audio?.input != null && Object.hasOwn(audio.input, 'transcription')) {
+      this.#inputTranscriptionEnabled = audio.input.transcription != null
+    } else if (Object.hasOwn(session, 'input_audio_transcription')) {
+      // Legacy flat field (older SDKs).
+      this.#inputTranscriptionEnabled = session.input_audio_transcription != null
+    }
+
+    inputFormat ??= session.input_audio_format
+    outputFormat ??= session.output_audio_format
+    voice ??= session.voice
+
+    if (inputFormat != null) {
+      this.#inputAudioMime = realtimeAudioFormatToMime(inputFormat)
+      this.#sessionConfig.input_audio_format = this.#inputAudioMime
+      const rate = toFiniteNumber(inputFormat.rate)
+      if (rate) this.#inputAudioRate = rate
+    }
+    if (outputFormat != null) {
+      this.#outputAudioMime = realtimeAudioFormatToMime(outputFormat)
+      this.#sessionConfig.output_audio_format = this.#outputAudioMime
+      const rate = toFiniteNumber(outputFormat.rate)
+      if (rate) this.#outputAudioRate = rate
+    }
+    if (voice != null) this.#sessionConfig.voice = String(voice)
+  }
+
+  /**
+   * @param {Record<string, unknown>} item
+   * @param {number} now
+   */
+  #absorbInputItem (item, now) {
+    if (item == null) return
+
+    // A tool result the app feeds back becomes a tool result on the next turn's input.
+    if (item.type === 'function_call_output') {
+      const toolId = String(item.call_id ?? '')
+      /** @type {ToolResult} */
+      const result = {
+        toolId,
+        result: item.output == null ? '' : String(item.output),
+        type: 'function_call_output',
+      }
+      // Label the result with the function name carried by the originating call.
+      const name = this.#toolCallNames.get(toolId)
+      if (name) {
+        result.name = name
+        this.#toolCallNames.delete(toolId)
+      }
+      this.#pendingInput.toolResults.push(result)
+      return
+    }
+
+    // Only user items contribute to the input turn; skip assistant and system items.
+    if (item.role != null && item.role !== 'user') return
+
+    const { content } = item
+    if (!content) return
+
+    for (const part of content) {
+      switch (part?.type) {
+        case 'input_text':
+        case 'text':
+          this.#pendingInput.text += part.text ?? ''
+          break
+        case 'input_audio':
+        case 'audio':
+          if (part.audio) {
+            this.#pendingInput.audio.append(part.audio, now, this.#inputAudioMime, this.#inputAudioRate)
+          }
+          if (part.transcript) this.#pendingInput.transcript += part.transcript
+      }
+    }
+  }
+}
+
+module.exports = RealtimeSession