@voicelayer/sdk 0.6.1 → 0.6.2

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/dist/index.js CHANGED
@@ -750,7 +750,7 @@ var init_consultation = __esm({
750
750
  });
751
751
  }
752
752
  });
753
- var CALL_CONTROL_PUBSUB_CHANNEL, CALL_CONTROL_ACK_CHANNEL, CALL_CONTROL_READY_KEY, CallEventKind, OP_CALL_EVENT_KIND, OpCallEventKind, commandId, SayCommand, HangupCommand, DtmfCommand, InstructCommand, InjectContextCommand, CallControlCommand;
753
+ var CALL_CONTROL_PUBSUB_CHANNEL, CALL_CONTROL_ACK_CHANNEL, CALL_CONTROL_READY_KEY, CallEventKind, OP_CALL_EVENT_KIND, OpCallEventKind, LatencyMs, TURN_LATENCY_MAX_PER_CALL, commandId, SayCommand, HangupCommand, DtmfCommand, InstructCommand, InjectContextCommand, CallControlCommand;
754
754
  var init_call_events = __esm({
755
755
  "../contracts/src/call-events.ts"() {
756
756
  init_consultation();
@@ -779,7 +779,9 @@ var init_call_events = __esm({
779
779
  "dtmf.sent",
780
780
  // A host's runtime intervention (burn-down G-6): call_say / call_send_guidance / call_inject_context / call_instruct,
781
781
  // with masked args and the API key — written by the API only, never through the worker's /events endpoint.
782
- "mcp.interaction"
782
+ "mcp.interaction",
783
+ // One per caller turn the agent answered (burn-down G-44): where that turn's wait went — see TurnLatencyPayload.
784
+ "turn.latency"
783
785
  ]);
784
786
  OP_CALL_EVENT_KIND = /^(tool|handoff|lookup|record|notify|engine)\.[a-z0-9_]{1,48}(\.[a-z0-9_]{1,48})?$/;
785
787
  OpCallEventKind = z.string().regex(OP_CALL_EVENT_KIND);
@@ -806,6 +808,18 @@ var init_call_events = __esm({
806
808
  code: z.number().int().min(0).max(15),
807
809
  participantId: z.string().optional()
808
810
  });
811
+ LatencyMs = z.number().int().min(0).max(6e5).nullable();
812
+ TURN_LATENCY_MAX_PER_CALL = 500;
813
+ z.object({
814
+ turn: z.number().int().min(1),
815
+ atMs: z.number().int().min(0),
816
+ endpointingMs: LatencyMs,
817
+ sttMs: LatencyMs,
818
+ llmMs: LatencyMs,
819
+ ttsMs: LatencyMs,
820
+ endToEndMs: LatencyMs,
821
+ realtime: z.boolean().optional()
822
+ });
809
823
  z.object({
810
824
  totalUsd: z.string(),
811
825
  breakdown: z.object({
@@ -924,9 +938,68 @@ var init_text_turns = __esm({
924
938
  ]);
925
939
  }
926
940
  });
941
+ var PLATFORM_MAX_CALL_MINUTES, SIP_LEG_MAX_CALL_DURATION_SECONDS, INBOUND_REFUSED_MESSAGE, Message, CallerRateLimit, CallLimits, CallerChannel, CallDurationLimits, INBOUND_REFUSED_END_REASON_PREFIX, CallAdmissionResponse;
942
+ var init_call_limits = __esm({
943
+ "../contracts/src/call-limits.ts"() {
944
+ PLATFORM_MAX_CALL_MINUTES = 240;
945
+ SIP_LEG_MAX_CALL_DURATION_SECONDS = (PLATFORM_MAX_CALL_MINUTES + 15) * 60;
946
+ INBOUND_REFUSED_MESSAGE = "Sorry, we can't take your call right now. Please try again later. Goodbye.";
947
+ Message = z.string().trim().min(1).max(300);
948
+ CallerRateLimit = z.object({
949
+ enabled: z.boolean().optional(),
950
+ perTenMinutes: z.number().int().min(1).max(100).optional(),
951
+ perDay: z.number().int().min(1).max(1e3).optional(),
952
+ withheldPerTenMinutes: z.number().int().min(1).max(100).optional(),
953
+ withheldPerDay: z.number().int().min(1).max(1e3).optional()
954
+ });
955
+ CallLimits = z.object({
956
+ /** The agent's own ceiling in minutes; clamped to the workspace's plan ceiling. Absent ⇒ DEFAULT_MAX_CALL_MINUTES. */
957
+ maxDurationMinutes: z.number().int().min(1).max(PLATFORM_MAX_CALL_MINUTES).optional(),
958
+ /** Seconds before the limit the wrap-up line is spoken; 0 ⇒ no warning. */
959
+ wrapUpWarningSeconds: z.number().int().min(0).max(300).optional(),
960
+ wrapUpMessage: Message.optional(),
961
+ goodbyeMessage: Message.optional(),
962
+ callerRateLimit: CallerRateLimit.optional(),
963
+ rateLimitedMessage: Message.optional()
964
+ });
965
+ CallerChannel = z.enum(["phone", "web"]);
966
+ z.object({
967
+ roomName: z.string().min(1).max(256),
968
+ /** From the dispatch metadata: lets the API admit (gate) a room whose call row doesn't exist yet. */
969
+ agentId: z.string().max(128).optional(),
970
+ phoneNumberId: z.string().max(128).optional(),
971
+ caller: z.object({
972
+ channel: CallerChannel,
973
+ /** The caller's number as the carrier presented it; null / absent / not E.164 ⇒ the withheld bucket. */
974
+ number: z.string().max(64).nullable().optional()
975
+ })
976
+ });
977
+ CallDurationLimits = z.object({
978
+ /** The agent's own ceiling (already clamped to the plan), or null when it set none. */
979
+ agentMaxSeconds: z.number().int().min(1).nullable(),
980
+ /** The plan ceiling — the platform bound no call outlives. */
981
+ ceilingSeconds: z.number().int().min(1),
982
+ wrapUpWarningSeconds: z.number().int().min(0),
983
+ wrapUpMessage: z.string(),
984
+ goodbyeMessage: z.string()
985
+ });
986
+ INBOUND_REFUSED_END_REASON_PREFIX = "inbound_refused:";
987
+ CallAdmissionResponse = z.discriminatedUnion("admitted", [
988
+ z.object({ admitted: z.literal(true), limits: CallDurationLimits }),
989
+ z.object({
990
+ admitted: z.literal(false),
991
+ /** 'rate_limited:caller' | 'inbound_refused:<code>' — already recorded as the call's end reason. */
992
+ endReason: z.string(),
993
+ /** The fixed line to speak before hanging up. */
994
+ message: z.string()
995
+ })
996
+ ]);
997
+ }
998
+ });
927
999
  var VoiceConfig, ModelConfig, SttConfig, RealtimeConfig, PipelineMode, ToolBinding, ProcessFieldType, ProcessFieldSpec, ProcessCompleteWhen, ProcessBackendAckSpec, ProcessSpec, TriggerConditionSpec, TriggerActionSpec, TriggerSpec, RequiredInfoField, AgentConfig;
928
1000
  var init_agent_config = __esm({
929
1001
  "../contracts/src/agent-config.ts"() {
1002
+ init_call_limits();
930
1003
  init_consultation();
931
1004
  VoiceConfig = z.object({
932
1005
  // Finite, vetted list — platform-level enum.
@@ -1090,6 +1163,9 @@ var init_agent_config = __esm({
1090
1163
  // LLM slot through the connector instead of a first-party model. Null/absent →
1091
1164
  // first-party pipeline (unchanged default). See docs/specs/voice-brain-connector.
1092
1165
  brainConnectorId: z.string().uuid().nullable().optional(),
1166
+ // Per-agent call bounds: caller rate limit, max duration and their spoken lines (burn-down G-41, G-43). Absent ⇒
1167
+ // every default (call-limits.ts) — the rate limit is ON by default.
1168
+ callLimits: CallLimits.optional(),
1093
1169
  // Provenance: 'sdk' (hand-coded defineAgent worker), 'flow' (canvas deploy),
1094
1170
  // 'playbook' (agent-mode deploy), or 'connector' (brain-connector agent).
1095
1171
  // DERIVED from the linked agents row's deploy tags (flowId / playbookId /
@@ -1445,6 +1521,7 @@ var init_agent_versions = __esm({
1445
1521
  "../contracts/src/agent-versions.ts"() {
1446
1522
  init_agent_config();
1447
1523
  init_agents();
1524
+ init_call_limits();
1448
1525
  init_consultation();
1449
1526
  z.object({
1450
1527
  systemPrompt: z.string().max(64e3),
@@ -1459,6 +1536,8 @@ var init_agent_versions = __esm({
1459
1536
  consultation: ConsultationPolicy,
1460
1537
  tags: z.array(z.string().max(64)).max(32).nullable(),
1461
1538
  brainConnectorId: z.string().uuid().nullable(),
1539
+ // Versions published before call limits existed carry none: read as null (every default).
1540
+ callLimits: CallLimits.nullable().default(null),
1462
1541
  flowVersion: z.number().int().nullable(),
1463
1542
  processSchema: ProcessSchemaDTO.nullable()
1464
1543
  });
@@ -1494,6 +1573,7 @@ var init_agent_versions = __esm({
1494
1573
  "language",
1495
1574
  "tools",
1496
1575
  "brainConnectorId",
1576
+ "callLimits",
1497
1577
  "consultation",
1498
1578
  "tags",
1499
1579
  "flowVersion",
@@ -1635,6 +1715,9 @@ var init_model_catalog = __esm({
1635
1715
  CostEstimateBilledPerMinute = z.object({
1636
1716
  platformFee: z.number().nonnegative(),
1637
1717
  providerPassthrough: z.number().nonnegative(),
1718
+ // The flat own-key fee per minute (burn-down B-26); 0 when every provider runs on the platform's keys. Optional
1719
+ // only for responses from an API older than the fee.
1720
+ ownKeyFee: z.number().nonnegative().optional(),
1638
1721
  telephony: z.number().nonnegative(),
1639
1722
  allIn: z.number().nonnegative()
1640
1723
  });
@@ -1900,9 +1983,46 @@ var init_call_review = __esm({
1900
1983
  });
1901
1984
  }
1902
1985
  });
1986
+ var CAPTURED_FIELDS_MAX, CAPTURED_STRING_MAX_CHARS, CAPTURED_ARRAY_MAX_ITEMS, CapturedFieldValue, CapturedField, StructuredOutputs;
1987
+ var init_call_outcome = __esm({
1988
+ "../contracts/src/call-outcome.ts"() {
1989
+ init_agent_config();
1990
+ CAPTURED_FIELDS_MAX = 200;
1991
+ CAPTURED_STRING_MAX_CHARS = 8e3;
1992
+ CAPTURED_ARRAY_MAX_ITEMS = 100;
1993
+ CapturedFieldValue = z.union([
1994
+ z.string().max(CAPTURED_STRING_MAX_CHARS),
1995
+ z.number().finite(),
1996
+ z.boolean(),
1997
+ z.array(z.string().max(CAPTURED_STRING_MAX_CHARS)).max(CAPTURED_ARRAY_MAX_ITEMS),
1998
+ z.null()
1999
+ ]);
2000
+ CapturedField = z.object({
2001
+ name: z.string().min(1).max(64),
2002
+ type: ProcessFieldType,
2003
+ required: z.boolean(),
2004
+ // false ⇒ the call never captured it; `value` is then null
2005
+ filled: z.boolean(),
2006
+ value: CapturedFieldValue
2007
+ });
2008
+ StructuredOutputs = z.record(CapturedFieldValue);
2009
+ z.object({
2010
+ // Every declared field, in declaration order; empty when the agent declares none.
2011
+ fields: z.array(CapturedField).max(CAPTURED_FIELDS_MAX)
2012
+ }).superRefine((body, ctx) => {
2013
+ const seen = /* @__PURE__ */ new Set();
2014
+ for (const f of body.fields) {
2015
+ if (seen.has(f.name)) ctx.addIssue({ code: z.ZodIssueCode.custom, message: `duplicate field ${f.name}` });
2016
+ seen.add(f.name);
2017
+ if (!f.filled && f.value !== null) ctx.addIssue({ code: z.ZodIssueCode.custom, message: `${f.name}: an unfilled field has no value` });
2018
+ }
2019
+ });
2020
+ }
2021
+ });
1903
2022
  var TranscriptTurn, ConsultationAudit, ToolCallAudit, McpInteractionAudit;
1904
2023
  var init_call_result = __esm({
1905
2024
  "../contracts/src/call-result.ts"() {
2025
+ init_call_outcome();
1906
2026
  TranscriptTurn = z.object({
1907
2027
  speaker: z.enum(["caller", "agent"]),
1908
2028
  text: z.string(),
@@ -1951,9 +2071,13 @@ var init_call_result = __esm({
1951
2071
  startedAt: z.string().datetime(),
1952
2072
  endedAt: z.string().datetime(),
1953
2073
  durationMs: z.number().int().min(0),
2074
+ // Final segments only, in call order.
1954
2075
  transcript: z.array(TranscriptTurn),
1955
- // Whatever the agent's process-schema captured. No prescribed shape.
1956
- structuredOutputs: z.record(z.unknown()).optional(),
2076
+ // The filled fields of the agent's process / flow, keyed by name (burn-down G-20). Absent when the agent declares no
2077
+ // fields, or its worker never reported them.
2078
+ structuredOutputs: StructuredOutputs.optional(),
2079
+ // Every declared field with what the call captured and whether it was filled — the same list `call.ended` carries.
2080
+ fields: z.array(CapturedField).optional(),
1957
2081
  recordingUrl: z.string().url().optional(),
1958
2082
  cost: z.object({
1959
2083
  totalUsd: z.string(),
@@ -1965,9 +2089,11 @@ var init_call_result = __esm({
1965
2089
  });
1966
2090
  }
1967
2091
  });
1968
- var WEBHOOK_EVENT_TYPES, CallWebhookDirection, CallWebhookOutcome, CallWebhookAttributes, Party;
2092
+ var WEBHOOK_EVENT_TYPES, CallWebhookDirection, CallWebhookOutcome, CallWebhookAttributes, Party, CALL_ENDED_PAYLOAD_VERSION, CallEndedTranscript;
1969
2093
  var init_webhook_events = __esm({
1970
2094
  "../contracts/src/webhook-events.ts"() {
2095
+ init_call_result();
2096
+ init_call_outcome();
1971
2097
  WEBHOOK_EVENT_TYPES = ["call.started", "call.ended", "recording.ready"];
1972
2098
  z.enum(WEBHOOK_EVENT_TYPES);
1973
2099
  CallWebhookDirection = z.enum(["inbound", "outbound", "web"]);
@@ -1985,7 +2111,18 @@ var init_webhook_events = __esm({
1985
2111
  // set, it is the same value `call.ended` carries; `call.ended` always has one.
1986
2112
  direction: CallWebhookDirection.nullable()
1987
2113
  });
2114
+ CALL_ENDED_PAYLOAD_VERSION = 2;
2115
+ CallEndedTranscript = z.object({
2116
+ // Final segments only (what was said, never an interim STT guess), in call order. Empty when `truncated`.
2117
+ turns: z.array(TranscriptTurn),
2118
+ // How many final turns the call has — the length of `turns` unless truncated.
2119
+ turnCount: z.number().int().min(0),
2120
+ truncated: z.boolean(),
2121
+ // Where the whole transcript is: GET /v1/calls/:callId/result (its `transcript`). Always set.
2122
+ fetchPath: z.string().regex(/^\/v1\/calls\/[0-9a-f-]{36}\/result$/)
2123
+ });
1988
2124
  z.object({
2125
+ version: z.literal(CALL_ENDED_PAYLOAD_VERSION),
1989
2126
  callId: z.string().uuid(),
1990
2127
  durationMs: z.number().int().min(0),
1991
2128
  endedAt: z.string().datetime(),
@@ -1996,7 +2133,20 @@ var init_webhook_events = __esm({
1996
2133
  toE164: Party,
1997
2134
  direction: CallWebhookDirection,
1998
2135
  outcome: CallWebhookOutcome,
1999
- attributes: CallWebhookAttributes
2136
+ attributes: CallWebhookAttributes,
2137
+ // Every field the agent declared (its process, or a flow's slots), in declaration order, with what the call captured
2138
+ // and whether it was filled. Empty when the agent declares none.
2139
+ fields: z.array(CapturedField),
2140
+ // The filled fields as one object keyed by name (as GET /v1/calls/:id/result returns them). Absent when the agent
2141
+ // declares no fields.
2142
+ structuredOutputs: StructuredOutputs.optional(),
2143
+ transcript: CallEndedTranscript,
2144
+ // true: the agent's worker confirmed its last transcript segment and captured field had landed before this payload
2145
+ // was built. false: built without that word (a worker that crashed or runs an SDK before 0.6.2, or one that didn't
2146
+ // answer within the hold) — the transcript and fields are what had arrived; GET /v1/calls/:id/result has any later.
2147
+ finalized: z.boolean(),
2148
+ // true: the payload would have passed CALL_ENDED_PAYLOAD_MAX_BYTES, so values were left out (see the cap) — fetch them.
2149
+ valuesOmitted: z.boolean()
2000
2150
  });
2001
2151
  z.object({
2002
2152
  callId: z.string().uuid(),
@@ -2942,166 +3092,6 @@ var init_telephony = __esm({
2942
3092
  });
2943
3093
  }
2944
3094
  });
2945
- var SHORTENER_HOSTS, PLACEHOLDER_HOSTS, PublicUrl, A2pBusinessType, A2pBusinessInfo, A2pOptInType, A2pUseCase, A2pMessagingProfile, A2pState;
2946
- var init_a2p = __esm({
2947
- "../contracts/src/a2p.ts"() {
2948
- SHORTENER_HOSTS = [
2949
- "bit.ly",
2950
- "tinyurl.com",
2951
- "t.co",
2952
- "goo.gl",
2953
- "ow.ly",
2954
- "is.gd",
2955
- "buff.ly",
2956
- "rebrand.ly",
2957
- "cutt.ly",
2958
- "shorturl.at",
2959
- "tiny.cc"
2960
- ];
2961
- PLACEHOLDER_HOSTS = ["acme.com", "example.com", "example.org", "test.com", "localhost"];
2962
- PublicUrl = z.string().url().refine((v) => v.startsWith("https://"), "must be https").refine((v) => {
2963
- try {
2964
- const host = new URL(v).hostname.toLowerCase().replace(/^www\./, "");
2965
- return !SHORTENER_HOSTS.includes(host);
2966
- } catch {
2967
- return false;
2968
- }
2969
- }, "public URL shorteners are rejected by carriers \u2014 use your own domain").refine((v) => {
2970
- try {
2971
- const host = new URL(v).hostname.toLowerCase().replace(/^www\./, "");
2972
- return !PLACEHOLDER_HOSTS.some((p) => host === p || host.endsWith(`.${p}`));
2973
- } catch {
2974
- return false;
2975
- }
2976
- }, "placeholder domain \u2014 reviewers will follow this link and reject the campaign");
2977
- A2pBusinessType = z.enum([
2978
- "Sole Proprietorship",
2979
- "Partnership",
2980
- "Corporation",
2981
- "Co-operative",
2982
- "Limited Liability Corporation",
2983
- "Non-profit Corporation"
2984
- ]);
2985
- A2pBusinessInfo = z.object({
2986
- legalName: z.string().trim().min(2).max(200),
2987
- /** EIN (US) or equivalent registration number. */
2988
- registrationNumber: z.string().trim().min(4).max(50),
2989
- businessType: A2pBusinessType,
2990
- /** Publicly reachable production site — not staging, not a 404. */
2991
- website: PublicUrl,
2992
- industry: z.string().trim().min(2).max(60),
2993
- address: z.object({
2994
- street: z.string().trim().min(2).max(200),
2995
- city: z.string().trim().min(1).max(100),
2996
- region: z.string().trim().min(1).max(100),
2997
- postalCode: z.string().trim().min(2).max(20),
2998
- isoCountry: z.string().trim().length(2)
2999
- }),
3000
- authorizedRep: z.object({
3001
- firstName: z.string().trim().min(1).max(100),
3002
- lastName: z.string().trim().min(1).max(100),
3003
- email: z.string().trim().email(),
3004
- phone: z.string().trim().regex(/^\+[1-9]\d{6,14}$/, "must be E.164"),
3005
- jobTitle: z.string().trim().min(2).max(100)
3006
- })
3007
- });
3008
- A2pOptInType = z.enum(["WEB_FORM", "PAPER_FORM", "VERBAL", "VIA_TEXT", "MOBILE_QR_CODE"]);
3009
- A2pUseCase = z.enum([
3010
- "MIXED",
3011
- "CUSTOMER_CARE",
3012
- "MARKETING",
3013
- "ACCOUNT_NOTIFICATION",
3014
- "2FA",
3015
- "DELIVERY_NOTIFICATION",
3016
- "HIGHER_EDUCATION",
3017
- "POLLING_VOTING",
3018
- "PUBLIC_SERVICE_ANNOUNCEMENT",
3019
- "LOW_VOLUME"
3020
- ]);
3021
- A2pMessagingProfile = z.object({
3022
- useCase: A2pUseCase,
3023
- /** Specific, not generic. "We send texts" gets rejected; describe the actual
3024
- * messages and when they are sent. */
3025
- description: z.string().trim().min(40).max(4096),
3026
- /**
3027
- * The single most-rejected field. Must describe HOW people opt in, state the
3028
- * message frequency, include the "message and data rates may apply"
3029
- * disclosure, and link to publicly reachable evidence. Twilio's API bounds it
3030
- * to 40–2049 characters.
3031
- */
3032
- messageFlow: z.string().trim().min(40).max(2049),
3033
- optInType: A2pOptInType,
3034
- /** Publicly accessible screenshots/pages showing the opt-in. Reviewers open
3035
- * these; anything behind a login fails. */
3036
- optInEvidenceUrls: z.array(PublicUrl).min(1).max(5),
3037
- /** Real messages the customer will send. Must reflect the declared use case
3038
- * and carry opt-out language. */
3039
- messageSamples: z.array(z.string().trim().min(10).max(1024)).min(2).max(5),
3040
- privacyPolicyUrl: PublicUrl,
3041
- termsAndConditionsUrl: PublicUrl,
3042
- hasEmbeddedLinks: z.boolean().default(false),
3043
- hasEmbeddedPhone: z.boolean().default(false)
3044
- });
3045
- z.object({
3046
- business: A2pBusinessInfo,
3047
- messaging: A2pMessagingProfile,
3048
- /** Register against Twilio's mock endpoints — exercises the full pipeline
3049
- * with no fees and no real carrier submission. Used in CI and staging. */
3050
- mock: z.boolean().default(false)
3051
- }).superRefine((v, ctx) => {
3052
- const flow = v.messaging.messageFlow.toLowerCase();
3053
- if (!/(msg|message)\s*(&|and)\s*data rates/.test(flow)) {
3054
- ctx.addIssue({
3055
- code: z.ZodIssueCode.custom,
3056
- path: ["messaging", "messageFlow"],
3057
- message: 'must include a "Message and data rates may apply" disclosure \u2014 carriers reject without it'
3058
- });
3059
- }
3060
- if (!/\d/.test(flow) || !/(msg|message|text)/.test(flow)) {
3061
- ctx.addIssue({
3062
- code: z.ZodIssueCode.custom,
3063
- path: ["messaging", "messageFlow"],
3064
- message: 'must state message frequency, e.g. "Up to 4 msgs/month"'
3065
- });
3066
- }
3067
- const hasOptOut = v.messaging.messageSamples.some((s) => /stop/i.test(s));
3068
- if (!hasOptOut) {
3069
- ctx.addIssue({
3070
- code: z.ZodIssueCode.custom,
3071
- path: ["messaging", "messageSamples"],
3072
- message: 'at least one sample must include opt-out language (e.g. "Reply STOP to opt out")'
3073
- });
3074
- }
3075
- });
3076
- A2pState = z.enum([
3077
- "none",
3078
- "profile_pending",
3079
- "profile_approved",
3080
- "profile_failed",
3081
- "brand_pending",
3082
- "brand_approved",
3083
- "brand_failed",
3084
- "campaign_pending",
3085
- "messaging_ready",
3086
- "campaign_failed"
3087
- ]);
3088
- z.object({
3089
- state: A2pState,
3090
- customerProfileSid: z.string().nullable(),
3091
- trustProductSid: z.string().nullable(),
3092
- brandSid: z.string().nullable(),
3093
- messagingServiceSid: z.string().nullable(),
3094
- campaignSid: z.string().nullable(),
3095
- mock: z.boolean(),
3096
- /** Carrier/Twilio rejection details, surfaced verbatim so the customer can
3097
- * fix the specific field rather than guess. */
3098
- failures: z.array(z.object({ code: z.number().nullable(), field: z.string().nullable(), message: z.string() })).default([]),
3099
- /** Plain-language next step for the dashboard. */
3100
- nextAction: z.string().nullable(),
3101
- updatedAt: z.string().nullable()
3102
- });
3103
- }
3104
- });
3105
3095
  var ConnectorMode, ConnectorStatus, ConnectorNormalize;
3106
3096
  var init_connector = __esm({
3107
3097
  "../contracts/src/connector.ts"() {
@@ -3218,7 +3208,7 @@ var init_connector_stream = __esm({
3218
3208
  ]);
3219
3209
  }
3220
3210
  });
3221
- var FLOW_BOOT_FAILURES;
3211
+ var FLOW_BOOT_FAILURES, AGENT_END_REASONS;
3222
3212
  var init_call_end_reason = __esm({
3223
3213
  "../contracts/src/call-end-reason.ts"() {
3224
3214
  FLOW_BOOT_FAILURES = [
@@ -3229,9 +3219,11 @@ var init_call_end_reason = __esm({
3229
3219
  "worker_identity_refused",
3230
3220
  "provider_unavailable"
3231
3221
  ];
3232
- z.enum(
3233
- FLOW_BOOT_FAILURES.map((cause) => `flow_boot:${cause}`)
3234
- );
3222
+ AGENT_END_REASONS = [
3223
+ "max_duration",
3224
+ ...FLOW_BOOT_FAILURES.map((cause) => `flow_boot:${cause}`)
3225
+ ];
3226
+ z.enum(AGENT_END_REASONS);
3235
3227
  }
3236
3228
  });
3237
3229
  var EnvironmentSpec;
@@ -4427,6 +4419,23 @@ var init_wallet = __esm({
4427
4419
  });
4428
4420
  }
4429
4421
  });
4422
+ var PipelineSlotKey;
4423
+ var init_pipeline_keys = __esm({
4424
+ "../contracts/src/pipeline-keys.ts"() {
4425
+ PipelineSlotKey = z.object({
4426
+ /** The provider as the slot reports it (LiveKit's label: `openai`, `api.openai.com`, `Deepgram`, …). */
4427
+ provider: z.string().min(1).max(200),
4428
+ /** true = the workspace's own key built it; false = the platform's. */
4429
+ byok: z.boolean()
4430
+ });
4431
+ z.object({
4432
+ llm: PipelineSlotKey.optional(),
4433
+ stt: PipelineSlotKey.optional(),
4434
+ tts: PipelineSlotKey.optional(),
4435
+ realtime: PipelineSlotKey.optional()
4436
+ });
4437
+ }
4438
+ });
4430
4439
 
4431
4440
  // ../contracts/src/index.ts
4432
4441
  var init_src = __esm({
@@ -4456,6 +4465,7 @@ var init_src = __esm({
4456
4465
  init_tests();
4457
4466
  init_call_review();
4458
4467
  init_call_result();
4468
+ init_call_outcome();
4459
4469
  init_webhook_events();
4460
4470
  init_numbers();
4461
4471
  init_project_limits();
@@ -4471,12 +4481,12 @@ var init_src = __esm({
4471
4481
  init_connection();
4472
4482
  init_secret_headers();
4473
4483
  init_telephony();
4474
- init_a2p();
4475
4484
  init_connector();
4476
4485
  init_connector_stream();
4477
4486
  init_flow_compile();
4478
4487
  init_flow_compile_check();
4479
4488
  init_call_end_reason();
4489
+ init_call_limits();
4480
4490
  init_flow_enrichment();
4481
4491
  init_environments();
4482
4492
  init_variables();
@@ -4492,6 +4502,7 @@ var init_src = __esm({
4492
4502
  init_wallet();
4493
4503
  init_network_address();
4494
4504
  init_dispatch_metadata();
4505
+ init_pipeline_keys();
4495
4506
  }
4496
4507
  });
4497
4508
 
@@ -11984,6 +11995,9 @@ var MemoryClient = class {
11984
11995
  );
11985
11996
  }
11986
11997
  };
11998
+
11999
+ // src/calls.ts
12000
+ init_src();
11987
12001
  var CallSchema = z.object({
11988
12002
  id: z.string(),
11989
12003
  projectId: z.string(),
@@ -12042,6 +12056,7 @@ z.object({
12042
12056
  });
12043
12057
  var CallStateSyncResponse = z.object({ ok: z.literal(true) }).or(z.object({}).passthrough());
12044
12058
  var DtmfEventResponse = z.object({ accepted: z.literal(true) }).or(z.object({}).passthrough());
12059
+ var CallFinalizeResponse = z.object({ ok: z.boolean() }).or(z.object({}).passthrough());
12045
12060
  var UsageReportResponse = z.object({ recorded: z.boolean() }).or(z.object({}).passthrough());
12046
12061
  var CallEventEnvelopeResponse = z.object({
12047
12062
  event: z.object({
@@ -12080,13 +12095,28 @@ var CallsClient = class {
12080
12095
  ...input.attributes !== void 0 ? { attributes: input.attributes } : {},
12081
12096
  ...input.participants !== void 0 ? { participants: input.participants } : {},
12082
12097
  ...input.legs !== void 0 ? { legs: input.legs } : {},
12083
- ...input.startedAt !== void 0 ? { startedAt: input.startedAt } : {}
12098
+ ...input.startedAt !== void 0 ? { startedAt: input.startedAt } : {},
12099
+ ...input.pipelineKeys !== void 0 ? { pipelineKeys: input.pipelineKeys } : {},
12100
+ ...input.finalizes !== void 0 ? { finalizes: input.finalizes } : {}
12084
12101
  }
12085
12102
  },
12086
12103
  CallByRoomResponse
12087
12104
  );
12088
12105
  return call;
12089
12106
  }
12107
+ /**
12108
+ * The call's admission (burn-down G-41/G-42/G-43), asked once right after the caller joins and before the agent
12109
+ * speaks: admitted with its duration limits, or refused with the end reason the platform recorded and the line to
12110
+ * speak before hanging up. Throws on a transport / server error — the runtime retries, then refuses (fail closed).
12111
+ */
12112
+ async admit(input) {
12113
+ return this.transport.request(
12114
+ // The runtime owns the retries (bounded, then a polite refusal): a server error comes straight back, and one
12115
+ // attempt never holds a caller longer than 2.5 s.
12116
+ { method: "POST", path: "/v1/calls/admission", body: input, noRetryStatuses: [500, 502, 503, 504], signal: AbortSignal.timeout(2500) },
12117
+ CallAdmissionResponse
12118
+ );
12119
+ }
12090
12120
  async syncState(callId, input) {
12091
12121
  await this.transport.request(
12092
12122
  {
@@ -12120,11 +12150,12 @@ var CallsClient = class {
12120
12150
  }
12121
12151
  // One-shot cascade usage report at session close — the platform prices it
12122
12152
  // into the call's bill (voice_components). Idempotent server-side.
12123
- async reportUsage(callId, report) {
12153
+ async reportUsage(callId, report, opts = {}) {
12124
12154
  await this.transport.request(
12125
12155
  {
12126
12156
  method: "POST",
12127
12157
  path: `/v1/calls/${encodeURIComponent(callId)}/usage-report`,
12158
+ ...opts.signal ? { signal: opts.signal } : {},
12128
12159
  body: {
12129
12160
  durationMs: report.durationMs,
12130
12161
  llm: report.llm,
@@ -12135,6 +12166,22 @@ var CallsClient = class {
12135
12166
  UsageReportResponse
12136
12167
  );
12137
12168
  }
12169
+ /**
12170
+ * The worker's last word on a call (burn-down G-20): what the call captured — every declared field — sent at the
12171
+ * session's close after the worker's other writes for the call have landed. The API keeps the fields on the call
12172
+ * and releases the call's held `call.ended` webhook. Idempotent.
12173
+ */
12174
+ async finalize(callId, input, opts = {}) {
12175
+ await this.transport.request(
12176
+ {
12177
+ method: "POST",
12178
+ path: `/v1/calls/${encodeURIComponent(callId)}/finalize`,
12179
+ ...opts.signal ? { signal: opts.signal } : {},
12180
+ body: { fields: input.fields }
12181
+ },
12182
+ CallFinalizeResponse
12183
+ );
12184
+ }
12138
12185
  async appendDtmfEvent(callId, input) {
12139
12186
  await this.transport.request(
12140
12187
  {
@@ -13304,6 +13351,38 @@ function buildRedisBrainPubSub(redisUrl) {
13304
13351
  // src/runtime/pipeline-from-config.ts
13305
13352
  init_providers2();
13306
13353
  init_registry();
13354
+
13355
+ // src/runtime/pipeline-key-owner.ts
13356
+ var vaultKeyed = /* @__PURE__ */ new WeakSet();
13357
+ var isObject = (v) => v !== null && (typeof v === "object" || typeof v === "function");
13358
+ function vaultKeyedFactory(factory) {
13359
+ return async (call) => {
13360
+ const provider = await factory(call);
13361
+ if (isObject(provider)) vaultKeyed.add(provider);
13362
+ return provider;
13363
+ };
13364
+ }
13365
+ function builtWithVaultKey(provider) {
13366
+ return isObject(provider) && vaultKeyed.has(provider);
13367
+ }
13368
+ function providerLabel(provider) {
13369
+ const label = isObject(provider) ? provider.provider : void 0;
13370
+ return typeof label === "string" && label.length > 0 ? label : "unknown";
13371
+ }
13372
+ function pipelineKeys(pipeline, envKeyIsCustomers) {
13373
+ const slot = (provider) => ({
13374
+ provider: providerLabel(provider),
13375
+ byok: builtWithVaultKey(provider) || envKeyIsCustomers
13376
+ });
13377
+ return {
13378
+ ...pipeline.llm != null ? { llm: slot(pipeline.llm) } : {},
13379
+ ...pipeline.stt != null ? { stt: slot(pipeline.stt) } : {},
13380
+ ...pipeline.tts != null ? { tts: slot(pipeline.tts) } : {},
13381
+ ...pipeline.realtime != null ? { realtime: slot(pipeline.realtime) } : {}
13382
+ };
13383
+ }
13384
+
13385
+ // src/runtime/pipeline-from-config.ts
13307
13386
  function modelsFromWireConfig(p, creds2) {
13308
13387
  const language = p.language ?? "en-US";
13309
13388
  if (p.mode === "realtime" && p.realtime) {
@@ -13317,17 +13396,20 @@ function modelsFromWireConfig(p, creds2) {
13317
13396
  out.tts = ttsFactory(p.voice.provider, p.voice.voiceId, p.voice.model, creds2?.[p.voice.provider]);
13318
13397
  return out;
13319
13398
  }
13399
+ function owned(factory, creds2) {
13400
+ return creds2 ? vaultKeyedFactory(factory) : factory;
13401
+ }
13320
13402
  function realtimeFactory(r, creds2) {
13321
- return resolveRealtime(r.provider, { model: r.model, ...r.voice ? { voice: r.voice } : {}, ...creds2 ? { creds: creds2 } : {} });
13403
+ return owned(resolveRealtime(r.provider, { model: r.model, ...r.voice ? { voice: r.voice } : {}, ...creds2 ? { creds: creds2 } : {} }), creds2);
13322
13404
  }
13323
13405
  function llmFactory(provider, model2, creds2) {
13324
- return resolveLlm(provider, { model: model2, ...creds2 ? { creds: creds2 } : {} });
13406
+ return owned(resolveLlm(provider, { model: model2, ...creds2 ? { creds: creds2 } : {} }), creds2);
13325
13407
  }
13326
13408
  function sttFactory(provider, model2, language, creds2) {
13327
- return resolveStt(provider, { ...model2 ? { model: model2 } : {}, language, ...creds2 ? { creds: creds2 } : {} });
13409
+ return owned(resolveStt(provider, { ...model2 ? { model: model2 } : {}, language, ...creds2 ? { creds: creds2 } : {} }), creds2);
13328
13410
  }
13329
13411
  function ttsFactory(provider, voiceId, model2, creds2) {
13330
- return resolveTts(provider, { voice: voiceId, ...model2 ? { model: model2 } : {}, ...creds2 ? { creds: creds2 } : {} });
13412
+ return owned(resolveTts(provider, { voice: voiceId, ...model2 ? { model: model2 } : {}, ...creds2 ? { creds: creds2 } : {} }), creds2);
13331
13413
  }
13332
13414
 
13333
13415
  // src/agent.ts
@@ -15513,6 +15595,9 @@ function extractText2(msg) {
15513
15595
  const trimmed = txt.trim();
15514
15596
  return trimmed.length === 0 ? null : trimmed;
15515
15597
  }
15598
+
15599
+ // src/runtime/handoff.ts
15600
+ init_src();
15516
15601
  function createHandoffRuntime(config, deps) {
15517
15602
  return {
15518
15603
  async handoff(to, opts) {
@@ -15661,7 +15746,9 @@ function buildDefaultSipOps(trunkOverride, fromNumber) {
15661
15746
  participantIdentity: identity,
15662
15747
  playDialtone: false,
15663
15748
  waitUntilAnswered: true,
15664
- ...fromNumber ? { fromNumber } : {}
15749
+ ...fromNumber ? { fromNumber } : {},
15750
+ // LiveKit's own ceiling on the transfer leg (burn-down G-43): it can't outlive the platform's longest call.
15751
+ maxCallDuration: SIP_LEG_MAX_CALL_DURATION_SECONDS
15665
15752
  });
15666
15753
  return { identity };
15667
15754
  }
@@ -16185,16 +16272,16 @@ async function exchangeWorkerToken(input) {
16185
16272
  const internalToken = process.env["INTERNAL_SERVICE_TOKEN"];
16186
16273
  if (!apiUrl || !internalToken) return null;
16187
16274
  const delays = input.retryDelaysMs ?? EXCHANGE_RETRY_DELAYS_MS;
16188
- const deadline = Date.now() + (input.budgetMs ?? EXCHANGE_BUDGET_MS);
16275
+ const deadline2 = Date.now() + (input.budgetMs ?? EXCHANGE_BUDGET_MS);
16189
16276
  for (let attempt = 0; ; attempt++) {
16190
- const left = deadline - Date.now();
16277
+ const left = deadline2 - Date.now();
16191
16278
  if (left <= 0) return null;
16192
16279
  const r = await exchangeOnce(input, apiUrl, internalToken, Math.min(input.attemptTimeoutMs ?? EXCHANGE_ATTEMPT_TIMEOUT_MS, left));
16193
16280
  if ("token" in r) return r.token;
16194
16281
  if ("stale" in r) throw new StaleRouteError(r.stale);
16195
16282
  const base = delays[attempt];
16196
16283
  const wait = base === void 0 ? void 0 : Math.max(base, "retryAfterMs" in r && r.retryAfterMs !== void 0 ? r.retryAfterMs : 0);
16197
- if (r.final || wait === void 0 || Date.now() + wait >= deadline) return null;
16284
+ if (r.final || wait === void 0 || Date.now() + wait >= deadline2) return null;
16198
16285
  await new Promise((resolve) => setTimeout(resolve, wait));
16199
16286
  }
16200
16287
  }
@@ -16393,7 +16480,7 @@ async function answerTextTurn(deps) {
16393
16480
  );
16394
16481
  const brain = pipeline?.brain === "bound";
16395
16482
  const label = pipeline?.llm ?? { provider: "openai", model: model2.model };
16396
- const byok = pipeline ? pipeline.byokProviders.includes(label.provider) : false;
16483
+ const byok = builtWithVaultKey(model2);
16397
16484
  const usage = brain ? [{ provider: "brain", model: model2.model || "brain", inputTokens: 0, outputTokens: 0, byok: true }] : [];
16398
16485
  const onMetrics = (m) => {
16399
16486
  if (brain) return;
@@ -17040,8 +17127,18 @@ function guardIsActive(opts) {
17040
17127
  return false;
17041
17128
  }
17042
17129
 
17130
+ // src/runtime/close-deadline.ts
17131
+ var CLOSE_REQUEST_TIMEOUT_MS = 5e3;
17132
+ function deadline(ms) {
17133
+ const controller = new AbortController();
17134
+ const timer = setTimeout(() => controller.abort(new Error(`timed out after ${ms} ms`)), ms);
17135
+ timer.unref?.();
17136
+ return { signal: controller.signal, clear: () => clearTimeout(timer) };
17137
+ }
17138
+
17043
17139
  // src/runtime/call-sync.ts
17044
17140
  var END_REASON_WAIT_MS = 2e3;
17141
+ var FINALIZE_WRITES_WAIT_MS = 3e3;
17045
17142
  function makeNoopHandle(callId) {
17046
17143
  return {
17047
17144
  callId,
@@ -17053,6 +17150,13 @@ function makeNoopHandle(callId) {
17053
17150
  publishDtmf() {
17054
17151
  },
17055
17152
  async recordEndReason() {
17153
+ },
17154
+ publishTurnLatency() {
17155
+ },
17156
+ track(write) {
17157
+ return write;
17158
+ },
17159
+ async finalize() {
17056
17160
  }
17057
17161
  };
17058
17162
  }
@@ -17065,6 +17169,16 @@ async function createCallSyncHandle(init) {
17065
17169
  });
17066
17170
  return makeNoopHandle(init.call.callId);
17067
17171
  }
17172
+ const inFlight = /* @__PURE__ */ new Set();
17173
+ const track = (write) => {
17174
+ const settled = write.then(
17175
+ () => void 0,
17176
+ () => void 0
17177
+ );
17178
+ inFlight.add(settled);
17179
+ void settled.then(() => inFlight.delete(settled));
17180
+ return write;
17181
+ };
17068
17182
  return {
17069
17183
  callId,
17070
17184
  resolved: true,
@@ -17074,13 +17188,13 @@ async function createCallSyncHandle(init) {
17074
17188
  startMs,
17075
17189
  (input.endAt ?? input.at).getTime() - init.call.startedAt.getTime()
17076
17190
  );
17077
- void init.client.calls.appendTranscriptSegment(callId, {
17191
+ void track(init.client.calls.appendTranscriptSegment(callId, {
17078
17192
  speaker: input.speaker,
17079
17193
  text: input.text,
17080
17194
  startMs,
17081
17195
  endMs,
17082
17196
  final: input.final
17083
- }).catch((err) => {
17197
+ })).catch((err) => {
17084
17198
  console.warn("[sdk.call-sync] segment publish failed", {
17085
17199
  speaker: input.speaker,
17086
17200
  err: err instanceof Error ? err.message : err
@@ -17088,12 +17202,43 @@ async function createCallSyncHandle(init) {
17088
17202
  });
17089
17203
  },
17090
17204
  syncState(input) {
17091
- void init.client.calls.syncState(callId, input).catch((err) => {
17205
+ void track(init.client.calls.syncState(callId, input)).catch((err) => {
17092
17206
  console.warn("[sdk.call-sync] state sync failed", {
17093
17207
  err: err instanceof Error ? err.message : err
17094
17208
  });
17095
17209
  });
17096
17210
  },
17211
+ publishTurnLatency(payload) {
17212
+ void track(init.client.calls.appendEvent(callId, { kind: "turn.latency", payload: { ...payload } })).catch((err) => {
17213
+ console.warn("[sdk.call-sync] turn latency publish failed", {
17214
+ turn: payload.turn,
17215
+ err: err instanceof Error ? err.message : err
17216
+ });
17217
+ });
17218
+ },
17219
+ track,
17220
+ async finalize(fields) {
17221
+ let timer;
17222
+ try {
17223
+ await Promise.race([
17224
+ Promise.all([...inFlight]),
17225
+ new Promise((resolve) => {
17226
+ timer = setTimeout(resolve, FINALIZE_WRITES_WAIT_MS);
17227
+ })
17228
+ ]);
17229
+ clearTimeout(timer);
17230
+ const limit = deadline(CLOSE_REQUEST_TIMEOUT_MS);
17231
+ try {
17232
+ await init.client.calls.finalize(callId, { fields: [...fields] }, { signal: limit.signal });
17233
+ } finally {
17234
+ limit.clear();
17235
+ }
17236
+ } catch (err) {
17237
+ console.warn("[sdk.call-sync] finalize failed", { err: err instanceof Error ? err.message : err });
17238
+ } finally {
17239
+ clearTimeout(timer);
17240
+ }
17241
+ },
17097
17242
  async recordEndReason(reason) {
17098
17243
  let timer;
17099
17244
  try {
@@ -17113,7 +17258,7 @@ async function createCallSyncHandle(init) {
17113
17258
  }
17114
17259
  },
17115
17260
  publishDtmf(input) {
17116
- void init.client.calls.appendDtmfEvent(callId, input).catch((err) => {
17261
+ void track(init.client.calls.appendDtmfEvent(callId, input)).catch((err) => {
17117
17262
  console.warn("[sdk.call-sync] dtmf publish failed", {
17118
17263
  direction: input.direction,
17119
17264
  digit: input.digit,
@@ -17138,7 +17283,10 @@ async function resolveCallIdWithRetry(init) {
17138
17283
  attributes: initialAttributes(init.call),
17139
17284
  participants: initialParticipants(init.call),
17140
17285
  legs: initialLegs(init.call, init.roomName),
17141
- startedAt: init.call.startedAt.toISOString()
17286
+ startedAt: init.call.startedAt.toISOString(),
17287
+ ...init.pipelineKeys ? { pipelineKeys: init.pipelineKeys } : {},
17288
+ // this worker sends POST /v1/calls/:id/finalize at the close: the API holds call.ended for it (G-20)
17289
+ finalizes: true
17142
17290
  });
17143
17291
  return summary.id;
17144
17292
  } catch {
@@ -17208,7 +17356,7 @@ function initialLegs(call, roomName) {
17208
17356
  startedAt
17209
17357
  }
17210
17358
  ];
17211
- if (destination) {
17359
+ if (destination && call.metadata["direction"] !== "outbound") {
17212
17360
  legs.unshift({
17213
17361
  legKey: "primary-inbound",
17214
17362
  kind: "sip-inbound",
@@ -17228,28 +17376,29 @@ function looksLikeSipParticipant(value) {
17228
17376
 
17229
17377
  // src/runtime/metrics-bridge.ts
17230
17378
  init_dist2();
17379
+ var payerKey = (byok) => byok === void 0 ? "" : byok ? "byok" : "platform";
17231
17380
  var UsageAccumulator = class {
17232
17381
  llm = /* @__PURE__ */ new Map();
17233
17382
  tts = /* @__PURE__ */ new Map();
17234
17383
  stt = /* @__PURE__ */ new Map();
17235
- // `byok` keeps a line whose payer is known (the SDK's helper calls) apart from LiveKit's own line on the same model,
17236
- // which the API prices by the workspace's vault keys.
17237
17384
  addLlm(provider, model2, inputTokens, outputTokens, byok) {
17238
- const key = `${provider}:${model2}:${byok === void 0 ? "" : byok ? "byok" : "platform"}`;
17385
+ const key = `${provider}:${model2}:${payerKey(byok)}`;
17239
17386
  const entry2 = this.llm.get(key) ?? { provider, model: model2, inputTokens: 0, outputTokens: 0, ...byok !== void 0 ? { byok } : {} };
17240
17387
  entry2.inputTokens += Math.max(0, inputTokens || 0);
17241
17388
  entry2.outputTokens += Math.max(0, outputTokens || 0);
17242
17389
  this.llm.set(key, entry2);
17243
17390
  }
17244
- addTts(provider, chars) {
17245
- const entry2 = this.tts.get(provider) ?? { provider, chars: 0 };
17391
+ addTts(provider, chars, byok) {
17392
+ const key = `${provider}:${payerKey(byok)}`;
17393
+ const entry2 = this.tts.get(key) ?? { provider, chars: 0, ...byok !== void 0 ? { byok } : {} };
17246
17394
  entry2.chars += Math.max(0, chars || 0);
17247
- this.tts.set(provider, entry2);
17395
+ this.tts.set(key, entry2);
17248
17396
  }
17249
- addStt(provider, seconds) {
17250
- const entry2 = this.stt.get(provider) ?? { provider, seconds: 0 };
17397
+ addStt(provider, seconds, byok) {
17398
+ const key = `${provider}:${payerKey(byok)}`;
17399
+ const entry2 = this.stt.get(key) ?? { provider, seconds: 0, ...byok !== void 0 ? { byok } : {} };
17251
17400
  entry2.seconds += Math.max(0, seconds || 0);
17252
- this.stt.set(provider, entry2);
17401
+ this.stt.set(key, entry2);
17253
17402
  }
17254
17403
  isEmpty() {
17255
17404
  return this.llm.size === 0 && this.tts.size === 0 && this.stt.size === 0;
@@ -17302,10 +17451,13 @@ function attachEndpointingProbe(session, events, ctx, deps) {
17302
17451
  });
17303
17452
  recordTurnLatencySpan({ ...base, stage: "endpointing", latencyMs: delayMs });
17304
17453
  recordTurnLatencySpan({ ...base, stage: "stt", latencyMs: delayMs });
17454
+ ctx.turns?.note({ stage: "endpointing", latencyMs: delayMs });
17455
+ ctx.turns?.note({ stage: "stt", latencyMs: delayMs });
17305
17456
  };
17306
17457
  deps.vad.onSpeechEnd((sample) => {
17307
17458
  if (!armed) return;
17308
17459
  speechEndAt = sample.speechEndAt;
17460
+ ctx.turns?.noteSpeechEnd(sample.speechEndAt);
17309
17461
  tryEmit();
17310
17462
  });
17311
17463
  session.on(events.userStateChanged, (event) => {
@@ -17347,7 +17499,8 @@ function dispatch(metrics2, ctx) {
17347
17499
  provider,
17348
17500
  ...model2 ? { model: model2 } : {},
17349
17501
  inputTokens: m.promptTokens,
17350
- outputTokens: m.completionTokens
17502
+ outputTokens: m.completionTokens,
17503
+ ...ctx.keyOwner?.llm ? { byok: ctx.keyOwner.llm.byok } : {}
17351
17504
  });
17352
17505
  recordLlmUsage({
17353
17506
  ...base,
@@ -17363,6 +17516,7 @@ function dispatch(metrics2, ctx) {
17363
17516
  } else if (typeof m.ttftMs === "number" && m.ttftMs >= 0) {
17364
17517
  recordTurnLatency({ ...base, firstTokenMs: m.ttftMs });
17365
17518
  recordTurnLatencySpan({ ...base, stage: "llm", latencyMs: m.ttftMs });
17519
+ ctx.turns?.note({ stage: "llm", latencyMs: m.ttftMs, startedAt: m.timestamp - Math.max(0, m.durationMs) });
17366
17520
  }
17367
17521
  break;
17368
17522
  }
@@ -17370,7 +17524,7 @@ function dispatch(metrics2, ctx) {
17370
17524
  const m = metrics2;
17371
17525
  const provider = m.metadata?.modelProvider ?? "unknown";
17372
17526
  const model2 = m.metadata?.modelName;
17373
- ctx.usage?.record({ kind: "tts", provider, chars: m.charactersCount });
17527
+ ctx.usage?.record({ kind: "tts", provider, chars: m.charactersCount, ...ctx.keyOwner?.tts ? { byok: ctx.keyOwner.tts.byok } : {} });
17374
17528
  recordTtsUsage({
17375
17529
  ...base,
17376
17530
  provider,
@@ -17380,6 +17534,7 @@ function dispatch(metrics2, ctx) {
17380
17534
  if (typeof m.ttfbMs === "number" && m.ttfbMs >= 0) {
17381
17535
  recordTurnLatency({ ...base, ttsStartMs: m.ttfbMs });
17382
17536
  recordTurnLatencySpan({ ...base, stage: "tts", latencyMs: m.ttfbMs });
17537
+ ctx.turns?.note({ stage: "tts", latencyMs: m.ttfbMs, startedAt: m.timestamp - Math.max(0, m.durationMs) });
17383
17538
  }
17384
17539
  break;
17385
17540
  }
@@ -17387,7 +17542,12 @@ function dispatch(metrics2, ctx) {
17387
17542
  const m = metrics2;
17388
17543
  const provider = m.metadata?.modelProvider ?? "unknown";
17389
17544
  const model2 = m.metadata?.modelName;
17390
- ctx.usage?.record({ kind: "stt", provider, seconds: (m.audioDurationMs ?? 0) / 1e3 });
17545
+ ctx.usage?.record({
17546
+ kind: "stt",
17547
+ provider,
17548
+ seconds: (m.audioDurationMs ?? 0) / 1e3,
17549
+ ...ctx.keyOwner?.stt ? { byok: ctx.keyOwner.stt.byok } : {}
17550
+ });
17391
17551
  recordSttUsage({
17392
17552
  ...base,
17393
17553
  provider,
@@ -17410,6 +17570,7 @@ function dispatch(metrics2, ctx) {
17410
17570
  if (typeof m.ttftMs === "number" && m.ttftMs >= 0) {
17411
17571
  recordTurnLatency({ ...base, firstTokenMs: m.ttftMs });
17412
17572
  recordTurnLatencySpan({ ...base, stage: "llm", latencyMs: m.ttftMs, realtime: true });
17573
+ ctx.turns?.note({ stage: "llm", latencyMs: m.ttftMs, startedAt: m.timestamp, realtime: true });
17413
17574
  }
17414
17575
  if (typeof m.sessionDurationMs === "number" && m.sessionDurationMs > 0) {
17415
17576
  recordRealtimeSessionDuration({
@@ -17431,6 +17592,11 @@ function dispatch(metrics2, ctx) {
17431
17592
  });
17432
17593
  recordTurnLatencySpan({ ...base, stage: "endpointing", latencyMs: m.endOfUtteranceDelayMs });
17433
17594
  recordTurnLatencySpan({ ...base, stage: "stt", latencyMs: m.transcriptionDelayMs });
17595
+ if (ctx.turns) {
17596
+ if (m.endOfUtteranceDelayMs > 0) ctx.turns.note({ stage: "endpointing", latencyMs: m.endOfUtteranceDelayMs });
17597
+ if (m.transcriptionDelayMs > 0) ctx.turns.note({ stage: "stt", latencyMs: m.transcriptionDelayMs });
17598
+ if (m.lastSpeakingTimeMs > 0) ctx.turns.noteSpeechEnd(m.lastSpeakingTimeMs);
17599
+ }
17434
17600
  break;
17435
17601
  }
17436
17602
  case "interruption_metrics": {
@@ -17468,10 +17634,10 @@ function createTelemetrySink(accumulator = new UsageAccumulator()) {
17468
17634
  );
17469
17635
  break;
17470
17636
  case "tts":
17471
- accumulator.addTts(sample.provider, sample.chars ?? 0);
17637
+ accumulator.addTts(sample.provider, sample.chars ?? 0, sample.byok);
17472
17638
  break;
17473
17639
  case "stt":
17474
- accumulator.addStt(sample.provider, sample.seconds ?? 0);
17640
+ accumulator.addStt(sample.provider, sample.seconds ?? 0, sample.byok);
17475
17641
  break;
17476
17642
  }
17477
17643
  },
@@ -17554,7 +17720,12 @@ async function reportUsageAtClose(opts) {
17554
17720
  try {
17555
17721
  await opts.meter.drain(opts.drainMs ?? HANGUP_DRAIN_MS);
17556
17722
  if (opts.usage.isEmpty()) return;
17557
- await opts.send(opts.usage.report(opts.durationMs));
17723
+ const limit = deadline(CLOSE_REQUEST_TIMEOUT_MS);
17724
+ try {
17725
+ await opts.send(opts.usage.report(opts.durationMs), limit.signal);
17726
+ } finally {
17727
+ limit.clear();
17728
+ }
17558
17729
  } catch (err) {
17559
17730
  opts.warn?.("[agent] usage report failed", { err: err instanceof Error ? err.message : String(err) });
17560
17731
  }
@@ -17603,6 +17774,7 @@ function attachTurnModelClock(session, events, ctx, deps = {}) {
17603
17774
  safely2(() => {
17604
17775
  recordTurnLatencySpan({ ...base, stage: "llm", latencyMs, allowZero: true });
17605
17776
  if (latencyMs > 0) recordTurnLatency({ ...base, firstTokenMs: latencyMs });
17777
+ deps.turns?.note({ stage: "llm", latencyMs, replyAt: t.windowEnd });
17606
17778
  });
17607
17779
  };
17608
17780
  const releaseParked = () => {
@@ -17719,6 +17891,145 @@ function attachTurnModelClock(session, events, ctx, deps = {}) {
17719
17891
  };
17720
17892
  }
17721
17893
 
17894
+ // src/runtime/turn-latency.ts
17895
+ init_src();
17896
+ var MAX_TURN_LATENCY_EVENTS = TURN_LATENCY_MAX_PER_CALL;
17897
+ var MAX_STAGE_MS = 6e5;
17898
+ var noStages = () => ({ endpointingMs: null, sttMs: null, llmMs: null, ttsMs: null, realtime: false });
17899
+ function attachTurnLatencyRecorder(session, events, deps) {
17900
+ const now = deps.now ?? Date.now;
17901
+ const maxTurns = deps.maxTurns ?? MAX_TURN_LATENCY_EVENTS;
17902
+ let waiting = null;
17903
+ let open = null;
17904
+ let turns = 0;
17905
+ let lastReplyAt = -Infinity;
17906
+ let lastSpeechEndAt = null;
17907
+ let closed = false;
17908
+ const ms = (v) => v === null || !Number.isFinite(v) || v < 0 || v > MAX_STAGE_MS ? null : Math.round(v);
17909
+ const emit = (t) => {
17910
+ const endpointingMs = ms(t.endpointingMs);
17911
+ const llmMs = ms(t.llmMs);
17912
+ const ttsMs = ms(t.ttsMs);
17913
+ const measured = t.speechEndAt !== null ? ms(t.replyAt - t.speechEndAt) : null;
17914
+ const summed = endpointingMs !== null && llmMs !== null && ttsMs !== null ? ms(endpointingMs + llmMs + ttsMs) : null;
17915
+ try {
17916
+ deps.emit({
17917
+ turn: t.index,
17918
+ atMs: Math.max(0, Math.round(t.replyAt - deps.callStartedAt)),
17919
+ endpointingMs,
17920
+ sttMs: ms(t.sttMs),
17921
+ llmMs,
17922
+ ttsMs,
17923
+ endToEndMs: t.realtime ? llmMs : measured ?? summed,
17924
+ ...t.realtime ? { realtime: true } : {}
17925
+ });
17926
+ } catch (err) {
17927
+ console.warn("[agent] turn latency emit failed", { err: err instanceof Error ? err.message : String(err) });
17928
+ }
17929
+ };
17930
+ const at = (event) => {
17931
+ const t = event?.createdAt;
17932
+ return typeof t === "number" ? t : now();
17933
+ };
17934
+ const targetFor = (s) => {
17935
+ if (waiting && (s.startedAt === void 0 || s.startedAt >= waiting.firstFinalAt)) return waiting;
17936
+ return open;
17937
+ };
17938
+ if (!deps.realtime) {
17939
+ session.on(events.userInputTranscribed, (event) => {
17940
+ if (closed || !event?.isFinal) return;
17941
+ waiting ??= { ...noStages(), firstFinalAt: at(event) };
17942
+ });
17943
+ session.on(events.agentStateChanged, (event) => {
17944
+ if (closed || event?.newState !== "speaking" || !waiting) return;
17945
+ const replyAt = at(event);
17946
+ if (open) emit(open);
17947
+ open = null;
17948
+ if (turns >= maxTurns) {
17949
+ waiting = null;
17950
+ return;
17951
+ }
17952
+ turns += 1;
17953
+ const speechEndAt = lastSpeechEndAt !== null && lastSpeechEndAt > lastReplyAt && lastSpeechEndAt <= replyAt ? lastSpeechEndAt : null;
17954
+ const { firstFinalAt: _first, ...stages } = waiting;
17955
+ open = { ...stages, index: turns, replyAt, speechEndAt };
17956
+ waiting = null;
17957
+ lastReplyAt = replyAt;
17958
+ });
17959
+ }
17960
+ return {
17961
+ note(s) {
17962
+ if (closed || !Number.isFinite(s.latencyMs) || s.latencyMs < 0) return;
17963
+ if (deps.realtime) {
17964
+ if (s.stage !== "llm" || turns >= maxTurns) return;
17965
+ turns += 1;
17966
+ emit({ ...noStages(), llmMs: s.latencyMs, realtime: true, index: turns, replyAt: (s.startedAt ?? now()) + s.latencyMs, speechEndAt: null });
17967
+ return;
17968
+ }
17969
+ switch (s.stage) {
17970
+ case "endpointing":
17971
+ case "stt": {
17972
+ const key = s.stage === "endpointing" ? "endpointingMs" : "sttMs";
17973
+ if (waiting) waiting[key] = s.latencyMs;
17974
+ else if (open && open[key] === null) open[key] = s.latencyMs;
17975
+ return;
17976
+ }
17977
+ case "llm": {
17978
+ if (s.replyAt !== void 0) {
17979
+ if (open && open.replyAt === s.replyAt) open.llmMs = (open.llmMs ?? 0) + s.latencyMs;
17980
+ return;
17981
+ }
17982
+ const t = targetFor(s);
17983
+ if (t) t.llmMs = (t.llmMs ?? 0) + s.latencyMs;
17984
+ return;
17985
+ }
17986
+ case "tts": {
17987
+ const t = targetFor(s);
17988
+ if (t && t.ttsMs === null) t.ttsMs = s.latencyMs;
17989
+ return;
17990
+ }
17991
+ }
17992
+ },
17993
+ noteSpeechEnd(at2) {
17994
+ if (Number.isFinite(at2) && at2 > 0) lastSpeechEndAt = at2;
17995
+ },
17996
+ close() {
17997
+ if (closed) return;
17998
+ closed = true;
17999
+ if (open) emit(open);
18000
+ open = null;
18001
+ waiting = null;
18002
+ }
18003
+ };
18004
+ }
18005
+
18006
+ // src/runtime/call-outcome.ts
18007
+ init_src();
18008
+ var cut = (s) => s.length > CAPTURED_STRING_MAX_CHARS ? s.slice(0, CAPTURED_STRING_MAX_CHARS) : s;
18009
+ function wireValue(value) {
18010
+ if (value === void 0 || value === null) return null;
18011
+ if (typeof value === "string") return value.trim().length === 0 ? null : cut(value);
18012
+ if (typeof value === "number") return Number.isFinite(value) ? value : null;
18013
+ if (typeof value === "boolean") return value;
18014
+ if (value instanceof Date) return Number.isNaN(value.getTime()) ? null : value.toISOString();
18015
+ if (Array.isArray(value)) {
18016
+ const items = value.filter((v) => v !== void 0 && v !== null).map((v) => cut(typeof v === "string" ? v : JSON.stringify(v))).slice(0, CAPTURED_ARRAY_MAX_ITEMS);
18017
+ return items.length === 0 ? null : items;
18018
+ }
18019
+ try {
18020
+ return cut(JSON.stringify(value));
18021
+ } catch {
18022
+ return null;
18023
+ }
18024
+ }
18025
+ function capturedFields(def, data) {
18026
+ if (!def) return [];
18027
+ return Object.entries(def.collect).filter(([name]) => name.length > 0 && name.length <= 64).slice(0, CAPTURED_FIELDS_MAX).map(([name, field]) => {
18028
+ const value = wireValue(data[name]);
18029
+ return { name, type: field.type, required: field.required === true, filled: value !== null, value };
18030
+ });
18031
+ }
18032
+
17722
18033
  // src/agent.ts
17723
18034
  init_model_meter();
17724
18035
 
@@ -17921,6 +18232,230 @@ function actionSpecToAction(spec) {
17921
18232
  if (spec.kind === "say") return { say: spec.text };
17922
18233
  return void 0;
17923
18234
  }
18235
+
18236
+ // src/runtime/call-admission.ts
18237
+ init_src();
18238
+
18239
+ // src/end-call.ts
18240
+ var defaultClock = {
18241
+ now: () => Date.now(),
18242
+ setTimeout: (fn, ms) => setTimeout(fn, ms),
18243
+ clearTimeout: (h) => clearTimeout(h)
18244
+ };
18245
+ var HOST_HANGUP_HOOK_LIMIT_MS = 2e3;
18246
+ var GUARANTEE_RETRY_MS = 1500;
18247
+ var GOODBYE_LIMIT_MS = 1e4;
18248
+ var DEFAULT_MAX_CALL_DURATION_MS = 36e5;
18249
+ function createEndCallController(deps) {
18250
+ const clock = deps.clock ?? defaultClock;
18251
+ const log = deps.log ?? (() => {
18252
+ });
18253
+ let ended = false;
18254
+ let idleTimer = null;
18255
+ let durationTimer = null;
18256
+ let warnTimer = null;
18257
+ const clearIdle = () => {
18258
+ if (idleTimer !== null) {
18259
+ clock.clearTimeout(idleTimer);
18260
+ idleTimer = null;
18261
+ }
18262
+ };
18263
+ const clearDuration = () => {
18264
+ if (durationTimer !== null) {
18265
+ clock.clearTimeout(durationTimer);
18266
+ durationTimer = null;
18267
+ }
18268
+ if (warnTimer !== null) {
18269
+ clock.clearTimeout(warnTimer);
18270
+ warnTimer = null;
18271
+ }
18272
+ };
18273
+ const sleep2 = (ms) => new Promise((resolve) => {
18274
+ clock.setTimeout(resolve, ms);
18275
+ });
18276
+ const guarantee = async (teardown = deps.runTeardown) => {
18277
+ if (deps.guaranteeMs === false) return;
18278
+ const deadline2 = clock.now() + deps.guaranteeMs;
18279
+ while (!deps.isEnded() && clock.now() < deadline2) {
18280
+ await sleep2(Math.min(GUARANTEE_RETRY_MS, Math.max(0, deadline2 - clock.now())));
18281
+ if (deps.isEnded()) return;
18282
+ log("[agent] end-call teardown not confirmed; retrying", { teardown: deps.teardown });
18283
+ try {
18284
+ await teardown();
18285
+ } catch (err) {
18286
+ log("[agent] end-call retry teardown threw", { err: String(err) });
18287
+ }
18288
+ }
18289
+ };
18290
+ let hostHangup = null;
18291
+ const hostHangUp = (reason) => {
18292
+ if (deps.isEnded()) return Promise.resolve();
18293
+ hostHangup ??= (async () => {
18294
+ ended = true;
18295
+ clearIdle();
18296
+ clearDuration();
18297
+ const info = { ...reason !== void 0 ? { reason } : {}, trigger: "host", teardown: "room" };
18298
+ deps.emit?.(info);
18299
+ log("[agent] hang up (host)", { reason });
18300
+ if (deps.onEnd) {
18301
+ let limit;
18302
+ await Promise.race([
18303
+ Promise.resolve().then(() => deps.onEnd(info)).catch((err) => log("[agent] endCall.onEnd threw on a host hangup", { err: String(err) })),
18304
+ new Promise((resolve) => {
18305
+ limit = clock.setTimeout(resolve, HOST_HANGUP_HOOK_LIMIT_MS);
18306
+ })
18307
+ ]).finally(() => clock.clearTimeout(limit));
18308
+ }
18309
+ const teardown = deps.runRoomTeardown ?? deps.runTeardown;
18310
+ try {
18311
+ await teardown();
18312
+ } catch (err) {
18313
+ log("[agent] host hangup teardown threw", { err: String(err) });
18314
+ }
18315
+ await guarantee(teardown);
18316
+ if (!deps.isEnded()) hostHangup = null;
18317
+ })();
18318
+ return hostHangup;
18319
+ };
18320
+ const hangUp = async (opts) => {
18321
+ if (opts?.trigger === "host") return hostHangUp(opts.reason);
18322
+ if (ended) return;
18323
+ ended = true;
18324
+ clearIdle();
18325
+ clearDuration();
18326
+ const info = {
18327
+ ...opts?.reason !== void 0 ? { reason: opts.reason } : {},
18328
+ trigger: opts?.trigger ?? "user",
18329
+ teardown: deps.teardown
18330
+ };
18331
+ deps.emit?.(info);
18332
+ log("[agent] hang up", { trigger: info.trigger, teardown: info.teardown });
18333
+ let suppress = false;
18334
+ if (deps.onEnd) {
18335
+ try {
18336
+ suppress = await deps.onEnd(info) === false;
18337
+ } catch (err) {
18338
+ log("[agent] endCall.onEnd threw; running default teardown", { err: String(err) });
18339
+ }
18340
+ }
18341
+ if (suppress) {
18342
+ log("[agent] endCall.onEnd suppressed default teardown", { trigger: info.trigger });
18343
+ return;
18344
+ }
18345
+ try {
18346
+ await deps.runTeardown();
18347
+ } catch (err) {
18348
+ log("[agent] end-call teardown threw", { err: String(err) });
18349
+ }
18350
+ await guarantee();
18351
+ };
18352
+ return {
18353
+ hangUp,
18354
+ armIdle() {
18355
+ if (ended || deps.idleHangupMs === false) return;
18356
+ clearIdle();
18357
+ idleTimer = clock.setTimeout(() => {
18358
+ idleTimer = null;
18359
+ void hangUp({ trigger: "idle", reason: "no activity on the line" });
18360
+ }, deps.idleHangupMs);
18361
+ },
18362
+ cancelIdle() {
18363
+ clearIdle();
18364
+ },
18365
+ armDuration() {
18366
+ if (ended || deps.maxCallDurationMs === false) return;
18367
+ clearDuration();
18368
+ const durationMs = deps.maxCallDurationMs ?? DEFAULT_MAX_CALL_DURATION_MS;
18369
+ const wrapUp = deps.wrapUp;
18370
+ if (wrapUp && wrapUp.warnBeforeMs > 0 && durationMs > wrapUp.warnBeforeMs) {
18371
+ warnTimer = clock.setTimeout(() => {
18372
+ warnTimer = null;
18373
+ if (ended) return;
18374
+ try {
18375
+ void Promise.resolve(wrapUp.warn()).catch((err) => log("[agent] wrap-up line failed", { err: String(err) }));
18376
+ } catch (err) {
18377
+ log("[agent] wrap-up line failed", { err: String(err) });
18378
+ }
18379
+ }, durationMs - wrapUp.warnBeforeMs);
18380
+ }
18381
+ durationTimer = clock.setTimeout(() => {
18382
+ durationTimer = null;
18383
+ void (async () => {
18384
+ if (wrapUp && !ended) {
18385
+ let limit;
18386
+ await Promise.race([
18387
+ wrapUp.goodbye().catch((err) => log("[agent] goodbye line failed", { err: String(err) })),
18388
+ new Promise((resolve) => {
18389
+ limit = clock.setTimeout(resolve, GOODBYE_LIMIT_MS);
18390
+ })
18391
+ ]).finally(() => clock.clearTimeout(limit));
18392
+ }
18393
+ await hangUp({
18394
+ trigger: "max_duration",
18395
+ reason: `call exceeded the ${durationMs}ms duration ceiling`
18396
+ });
18397
+ })();
18398
+ }, durationMs);
18399
+ },
18400
+ noteEngineClosed() {
18401
+ if (ended) return;
18402
+ log("[agent] realtime engine closed; guaranteeing teardown");
18403
+ void hangUp({ trigger: "engine-closed", reason: "realtime engine closed" });
18404
+ },
18405
+ dispose() {
18406
+ clearIdle();
18407
+ clearDuration();
18408
+ }
18409
+ };
18410
+ }
18411
+
18412
+ // src/runtime/call-admission.ts
18413
+ var CALL_ADMISSION_RETRY_DELAYS_MS = [300, 1e3];
18414
+ var ADMISSION_UNAVAILABLE_END_REASON = `${INBOUND_REFUSED_END_REASON_PREFIX}gate_unavailable`;
18415
+ function notAVerdict(err) {
18416
+ if (err instanceof VoiceLayerAuthError || err instanceof VoiceLayerValidationError) return true;
18417
+ return err instanceof VoiceLayerHttpError && err.status >= 400 && err.status < 500 && err.status !== 408 && err.status !== 429;
18418
+ }
18419
+ async function requestCallAdmission(admit, req, opts = {}) {
18420
+ if (!admit) return { kind: "admitted", limits: null };
18421
+ const delays = opts.delaysMs ?? CALL_ADMISSION_RETRY_DELAYS_MS;
18422
+ const sleep2 = opts.sleep ?? ((ms) => new Promise((r) => setTimeout(r, ms)));
18423
+ const log = opts.log ?? ((msg, attrs2) => console.warn(msg, attrs2 ?? {}));
18424
+ for (let attempt = 1; ; attempt++) {
18425
+ try {
18426
+ const res = await admit(req);
18427
+ return res.admitted ? { kind: "admitted", limits: res.limits } : { kind: "refused", endReason: res.endReason, message: res.message };
18428
+ } catch (err) {
18429
+ if (notAVerdict(err)) {
18430
+ if (opts.platformWorker) {
18431
+ log("[agent] call admission refused for this room on a platform worker; refusing the call (fail closed)", { err: String(err) });
18432
+ return { kind: "refused", endReason: ADMISSION_UNAVAILABLE_END_REASON, message: INBOUND_REFUSED_MESSAGE };
18433
+ }
18434
+ log("[agent] call admission unavailable for this room; running on the agent settings", { err: String(err) });
18435
+ return { kind: "admitted", limits: null };
18436
+ }
18437
+ if (attempt > delays.length) {
18438
+ log("[agent] call admission failed after retries; refusing the call (fail closed)", { err: String(err), attempts: attempt });
18439
+ return { kind: "refused", endReason: ADMISSION_UNAVAILABLE_END_REASON, message: INBOUND_REFUSED_MESSAGE };
18440
+ }
18441
+ await sleep2(delays[attempt - 1]);
18442
+ }
18443
+ }
18444
+ }
18445
+ function effectiveMaxCallDurationMs(code, limits) {
18446
+ if (!limits) return code === void 0 ? DEFAULT_MAX_CALL_DURATION_MS : code;
18447
+ const ceiling = limits.ceilingSeconds * 1e3;
18448
+ const base = limits.agentMaxSeconds !== null ? limits.agentMaxSeconds * 1e3 : code === void 0 ? DEFAULT_MAX_CALL_DURATION_MS : code;
18449
+ return base === false ? ceiling : Math.min(base, ceiling);
18450
+ }
18451
+ async function playAdmissionRefusal(verdict, deps) {
18452
+ try {
18453
+ await deps.say(verdict.message, { allowInterruptions: false });
18454
+ } catch (err) {
18455
+ (deps.log ?? ((m, a) => console.warn(m, a ?? {})))("[agent] admission refusal line failed to play", { err: String(err) });
18456
+ }
18457
+ await deps.hangUp(verdict.endReason);
18458
+ }
17924
18459
  init_src();
17925
18460
  var AskHostInput = z.object({
17926
18461
  question: z.string().min(1).max(2e3),
@@ -18471,7 +19006,7 @@ function asText(value) {
18471
19006
  function masked(value) {
18472
19007
  return maskPii(asText(value).slice(0, READ_CAP));
18473
19008
  }
18474
- var cut = (text, cap) => text.length > cap ? `${text.slice(0, cap)}\u2026` : text;
19009
+ var cut2 = (text, cap) => text.length > cap ? `${text.slice(0, cap)}\u2026` : text;
18475
19010
  function toolAuditPayload(r) {
18476
19011
  const args = masked(r.params);
18477
19012
  const result = r.ok ? masked(r.result) : void 0;
@@ -18480,11 +19015,11 @@ function toolAuditPayload(r) {
18480
19015
  const payload = {
18481
19016
  invocationId: r.invocationId,
18482
19017
  toolName: r.toolName,
18483
- args: cut(args, cap),
18484
- ...result !== void 0 ? { result: cut(result, cap) } : {},
19018
+ args: cut2(args, cap),
19019
+ ...result !== void 0 ? { result: cut2(result, cap) } : {},
18485
19020
  ok: r.ok,
18486
19021
  blocked: r.blocked,
18487
- ...error !== void 0 ? { errorMessage: cut(error, Math.min(cap, 300)) } : {},
19022
+ ...error !== void 0 ? { errorMessage: cut2(error, Math.min(cap, 300)) } : {},
18488
19023
  startedAt: r.startedAt.toISOString(),
18489
19024
  durationMs: r.durationMs,
18490
19025
  // the tool got the caller's values in place of the call's PII tokens (G-33): how many and where — never a value
@@ -18504,150 +19039,6 @@ function toolAuditObserver(emit) {
18504
19039
  };
18505
19040
  }
18506
19041
 
18507
- // src/end-call.ts
18508
- var defaultClock = {
18509
- now: () => Date.now(),
18510
- setTimeout: (fn, ms) => setTimeout(fn, ms),
18511
- clearTimeout: (h) => clearTimeout(h)
18512
- };
18513
- var HOST_HANGUP_HOOK_LIMIT_MS = 2e3;
18514
- var GUARANTEE_RETRY_MS = 1500;
18515
- var DEFAULT_MAX_CALL_DURATION_MS = 36e5;
18516
- function createEndCallController(deps) {
18517
- const clock = deps.clock ?? defaultClock;
18518
- const log = deps.log ?? (() => {
18519
- });
18520
- let ended = false;
18521
- let idleTimer = null;
18522
- let durationTimer = null;
18523
- const clearIdle = () => {
18524
- if (idleTimer !== null) {
18525
- clock.clearTimeout(idleTimer);
18526
- idleTimer = null;
18527
- }
18528
- };
18529
- const clearDuration = () => {
18530
- if (durationTimer !== null) {
18531
- clock.clearTimeout(durationTimer);
18532
- durationTimer = null;
18533
- }
18534
- };
18535
- const sleep2 = (ms) => new Promise((resolve) => {
18536
- clock.setTimeout(resolve, ms);
18537
- });
18538
- const guarantee = async (teardown = deps.runTeardown) => {
18539
- if (deps.guaranteeMs === false) return;
18540
- const deadline = clock.now() + deps.guaranteeMs;
18541
- while (!deps.isEnded() && clock.now() < deadline) {
18542
- await sleep2(Math.min(GUARANTEE_RETRY_MS, Math.max(0, deadline - clock.now())));
18543
- if (deps.isEnded()) return;
18544
- log("[agent] end-call teardown not confirmed; retrying", { teardown: deps.teardown });
18545
- try {
18546
- await teardown();
18547
- } catch (err) {
18548
- log("[agent] end-call retry teardown threw", { err: String(err) });
18549
- }
18550
- }
18551
- };
18552
- let hostHangup = null;
18553
- const hostHangUp = (reason) => {
18554
- if (deps.isEnded()) return Promise.resolve();
18555
- hostHangup ??= (async () => {
18556
- ended = true;
18557
- clearIdle();
18558
- clearDuration();
18559
- const info = { ...reason !== void 0 ? { reason } : {}, trigger: "host", teardown: "room" };
18560
- deps.emit?.(info);
18561
- log("[agent] hang up (host)", { reason });
18562
- if (deps.onEnd) {
18563
- let limit;
18564
- await Promise.race([
18565
- Promise.resolve().then(() => deps.onEnd(info)).catch((err) => log("[agent] endCall.onEnd threw on a host hangup", { err: String(err) })),
18566
- new Promise((resolve) => {
18567
- limit = clock.setTimeout(resolve, HOST_HANGUP_HOOK_LIMIT_MS);
18568
- })
18569
- ]).finally(() => clock.clearTimeout(limit));
18570
- }
18571
- const teardown = deps.runRoomTeardown ?? deps.runTeardown;
18572
- try {
18573
- await teardown();
18574
- } catch (err) {
18575
- log("[agent] host hangup teardown threw", { err: String(err) });
18576
- }
18577
- await guarantee(teardown);
18578
- if (!deps.isEnded()) hostHangup = null;
18579
- })();
18580
- return hostHangup;
18581
- };
18582
- const hangUp = async (opts) => {
18583
- if (opts?.trigger === "host") return hostHangUp(opts.reason);
18584
- if (ended) return;
18585
- ended = true;
18586
- clearIdle();
18587
- clearDuration();
18588
- const info = {
18589
- ...opts?.reason !== void 0 ? { reason: opts.reason } : {},
18590
- trigger: opts?.trigger ?? "user",
18591
- teardown: deps.teardown
18592
- };
18593
- deps.emit?.(info);
18594
- log("[agent] hang up", { trigger: info.trigger, teardown: info.teardown });
18595
- let suppress = false;
18596
- if (deps.onEnd) {
18597
- try {
18598
- suppress = await deps.onEnd(info) === false;
18599
- } catch (err) {
18600
- log("[agent] endCall.onEnd threw; running default teardown", { err: String(err) });
18601
- }
18602
- }
18603
- if (suppress) {
18604
- log("[agent] endCall.onEnd suppressed default teardown", { trigger: info.trigger });
18605
- return;
18606
- }
18607
- try {
18608
- await deps.runTeardown();
18609
- } catch (err) {
18610
- log("[agent] end-call teardown threw", { err: String(err) });
18611
- }
18612
- await guarantee();
18613
- };
18614
- return {
18615
- hangUp,
18616
- armIdle() {
18617
- if (ended || deps.idleHangupMs === false) return;
18618
- clearIdle();
18619
- idleTimer = clock.setTimeout(() => {
18620
- idleTimer = null;
18621
- void hangUp({ trigger: "idle", reason: "no activity on the line" });
18622
- }, deps.idleHangupMs);
18623
- },
18624
- cancelIdle() {
18625
- clearIdle();
18626
- },
18627
- armDuration() {
18628
- if (ended || deps.maxCallDurationMs === false) return;
18629
- clearDuration();
18630
- const durationMs = deps.maxCallDurationMs ?? DEFAULT_MAX_CALL_DURATION_MS;
18631
- durationTimer = clock.setTimeout(() => {
18632
- durationTimer = null;
18633
- void hangUp({
18634
- trigger: "max_duration",
18635
- reason: `call exceeded the ${durationMs}ms duration ceiling`
18636
- });
18637
- }, durationMs);
18638
- },
18639
- noteEngineClosed() {
18640
- if (ended) return;
18641
- log("[agent] realtime engine closed; guaranteeing teardown");
18642
- void hangUp({ trigger: "engine-closed", reason: "realtime engine closed" });
18643
- },
18644
- dispose() {
18645
- clearIdle();
18646
- clearDuration();
18647
- }
18648
- };
18649
- }
18650
-
18651
19042
  // src/agent.ts
18652
19043
  init_helper_models();
18653
19044
  var FLOW_HANGUP_GRACE_MS = Number(process.env["VL_FLOW_HANGUP_GRACE_MS"] ?? "1200");
@@ -18962,6 +19353,7 @@ var Agent = class {
18962
19353
  const built = await resolveCallPipeline(effectiveConfig.models ?? {}, callInfo, job.proc.userData.vad);
18963
19354
  const pipeline = built.pipeline;
18964
19355
  if (built.failure && !flowBootFailure) flowBootFailure = built.failure;
19356
+ const keyOwner = pipelineKeys(pipeline, callCredential?.source !== "worker-token");
18965
19357
  const vadTap = tapVad(pipeline.vad, { endOfSpeechType: VADEventType.END_OF_SPEECH });
18966
19358
  const usingRealtime = pipeline.realtime != null;
18967
19359
  let runtimeLlm = usingRealtime ? pipeline.realtime : pipeline.llm;
@@ -18974,9 +19366,10 @@ var Agent = class {
18974
19366
  seedCallPrefill(processRt, dispatchMdEarly);
18975
19367
  await seedProjectVariables(processRt, callCredential);
18976
19368
  let opEventCallId = null;
19369
+ let trackWrite = (write) => write;
18977
19370
  const emitOpEvent = (kind, payload) => {
18978
19371
  if (!sdkClient || !opEventCallId) return;
18979
- void sdkClient.calls.appendEvent(opEventCallId, { kind, payload }).catch(() => {
19372
+ void trackWrite(sdkClient.calls.appendEvent(opEventCallId, { kind, payload })).catch(() => {
18980
19373
  });
18981
19374
  };
18982
19375
  const graphEvents = graphMode ? withOpEventFanout(createGraphEvents(processRt), emitOpEvent) : null;
@@ -19003,8 +19396,23 @@ var Agent = class {
19003
19396
  const callSync = await createCallSyncHandle({
19004
19397
  client: sdkClient,
19005
19398
  roomName: job.room.name ?? `room-${job.job?.id ?? "unknown"}`,
19006
- call: callInfo
19399
+ call: callInfo,
19400
+ pipelineKeys: keyOwner
19007
19401
  });
19402
+ trackWrite = callSync.track;
19403
+ const callerIsPhone = isSipParticipant2(participant);
19404
+ const callAdmission = await requestCallAdmission(
19405
+ sdkClient ? (req) => sdkClient.calls.admit(req) : null,
19406
+ {
19407
+ roomName: job.room.name ?? `room-${job.job?.id ?? "unknown"}`,
19408
+ ...stringValue(callInfo.metadata["agentId"]) ? { agentId: stringValue(callInfo.metadata["agentId"]) } : {},
19409
+ ...stringValue(callInfo.metadata["phoneNumberId"]) ? { phoneNumberId: stringValue(callInfo.metadata["phoneNumberId"]) } : {},
19410
+ caller: { channel: callerIsPhone ? "phone" : "web", number: callerIsPhone ? callInfo.callerId : null }
19411
+ },
19412
+ // our pool's per-call worker token: no "no verdict" — a missing row is a refusal, never an unbilled call
19413
+ { platformWorker: callCredential?.source === "worker-token" }
19414
+ );
19415
+ const admissionLimits = callAdmission.kind === "admitted" ? callAdmission.limits : null;
19008
19416
  const transcriptLines = [];
19009
19417
  const transcript = createTranscriptHandle(() => transcriptLines);
19010
19418
  const ctxRef = { current: null };
@@ -19136,7 +19544,8 @@ ${callIntent}`
19136
19544
  const endCallTeardown = endCallPolicy.teardown ?? "room";
19137
19545
  const endCallGuaranteeMs = endCallPolicy.guaranteeMs === void 0 ? 4e3 : endCallPolicy.guaranteeMs;
19138
19546
  const endCallIdleMs = endCallPolicy.idleHangupMs === void 0 ? 2e4 : endCallPolicy.idleHangupMs;
19139
- const endCallMaxDurationMs = endCallPolicy.maxCallDurationMs === void 0 ? DEFAULT_MAX_CALL_DURATION_MS : endCallPolicy.maxCallDurationMs;
19547
+ const endCallMaxDurationMs = effectiveMaxCallDurationMs(endCallPolicy.maxCallDurationMs, admissionLimits);
19548
+ const lineRef = { say: null };
19140
19549
  const endCallRoomName = job.room.name ?? `room-${job.job?.id ?? "unknown"}`;
19141
19550
  const localDisconnect = async () => {
19142
19551
  try {
@@ -19149,6 +19558,17 @@ ${callIntent}`
19149
19558
  guaranteeMs: endCallGuaranteeMs,
19150
19559
  idleHangupMs: endCallIdleMs,
19151
19560
  maxCallDurationMs: endCallMaxDurationMs,
19561
+ ...admissionLimits ? {
19562
+ wrapUp: {
19563
+ warnBeforeMs: admissionLimits.wrapUpWarningSeconds * 1e3,
19564
+ warn: () => lineRef.say?.(admissionLimits.wrapUpMessage),
19565
+ goodbye: async () => {
19566
+ const reported = callSync.recordEndReason("max_duration");
19567
+ await lineRef.say?.(admissionLimits.goodbyeMessage);
19568
+ await reported;
19569
+ }
19570
+ }
19571
+ } : {},
19152
19572
  // 'room' deletes the whole LK room (drops the SIP/PSTN leg → carrier BYE);
19153
19573
  // 'agent' only disconnects the agent participant.
19154
19574
  runTeardown: endCallTeardown === "room" ? () => endRoomBestEffort(endCallRoomName, localDisconnect) : localDisconnect,
@@ -19187,6 +19607,7 @@ ${callIntent}`
19187
19607
  }
19188
19608
  });
19189
19609
  const session = adapter.raw.session;
19610
+ lineRef.say = (text) => adapter.say(text, { allowInterruptions: false });
19190
19611
  sessionRef.current = session;
19191
19612
  const handoffRt = createHandoffRuntime(effectiveConfig.handoff, {
19192
19613
  session,
@@ -19423,6 +19844,7 @@ ${callIntent}`
19423
19844
  })
19424
19845
  });
19425
19846
  let usageReported = Promise.resolve();
19847
+ let finalized = Promise.resolve();
19426
19848
  session.once(Events.Close, () => {
19427
19849
  sessionClosed = true;
19428
19850
  endCallController.dispose();
@@ -19440,7 +19862,7 @@ ${callIntent}`
19440
19862
  meter: modelMeter,
19441
19863
  usage: usageSink,
19442
19864
  durationMs: outcome.durationMs,
19443
- send: (report) => calls.reportUsage(callSync.callId, report),
19865
+ send: (report, signal) => calls.reportUsage(callSync.callId, report, { signal }),
19444
19866
  warn: (message, meta) => console.warn(message, { callId: callSync.callId, ...meta })
19445
19867
  });
19446
19868
  }
@@ -19580,6 +20002,11 @@ ${callIntent}`
19580
20002
  callId: callSync.callId,
19581
20003
  ...callCampaignId ? { campaignId: callCampaignId } : {}
19582
20004
  };
20005
+ const turnLatency = attachTurnLatencyRecorder(
20006
+ session,
20007
+ { userInputTranscribed: Events.UserInputTranscribed, agentStateChanged: Events.AgentStateChanged },
20008
+ { callStartedAt: callInfo.startedAt.getTime(), emit: (payload) => callSync.publishTurnLatency(payload), realtime: usingRealtime }
20009
+ );
19583
20010
  const modelClock = graphMode && !usingRealtime ? attachTurnModelClock(
19584
20011
  session,
19585
20012
  {
@@ -19588,13 +20015,26 @@ ${callIntent}`
19588
20015
  agentStateChanged: Events.AgentStateChanged,
19589
20016
  speechCreated: Events.SpeechCreated
19590
20017
  },
19591
- metricsCtx
20018
+ metricsCtx,
20019
+ { turns: turnLatency }
19592
20020
  ) : void 0;
19593
20021
  if (modelClock) session.once(Events.Close, () => modelClock.close());
20022
+ session.once(Events.Close, () => {
20023
+ turnLatency.close();
20024
+ finalized = (async () => {
20025
+ await modelMeter.drain(HANGUP_DRAIN_MS);
20026
+ await new Promise((resolve) => setImmediate(resolve));
20027
+ await callSync.finalize(capturedFields(effectiveConfig.process, processRt.getData()));
20028
+ })().catch((err) => {
20029
+ console.warn("[agent] finalize failed", { callId: callSync.callId, err: err instanceof Error ? err.message : String(err) });
20030
+ });
20031
+ });
19594
20032
  modelMeter.bind(metricsCtx, modelClock);
19595
20033
  attachMetricsBridge(session, Events.MetricsCollected, {
19596
20034
  ...metricsCtx,
19597
20035
  usage: usageSink,
20036
+ keyOwner,
20037
+ turns: turnLatency,
19598
20038
  ...modelClock ? { modelClock } : {}
19599
20039
  });
19600
20040
  if (graphMode) {
@@ -19604,7 +20044,7 @@ ${callIntent}`
19604
20044
  userStateChanged: Events.UserStateChanged,
19605
20045
  userInputTranscribed: Events.UserInputTranscribed
19606
20046
  },
19607
- metricsCtx,
20047
+ { ...metricsCtx, turns: turnLatency },
19608
20048
  { vad: vadTap }
19609
20049
  );
19610
20050
  }
@@ -19667,11 +20107,11 @@ ${callIntent}`
19667
20107
  room: job.room.name ?? null,
19668
20108
  participant: participant.identity
19669
20109
  });
19670
- endCallController.armDuration();
20110
+ if (callAdmission.kind !== "refused") endCallController.armDuration();
19671
20111
  {
19672
20112
  const announceMd = parseJobMetadata(job.job?.metadata);
19673
20113
  const announcement = stringValue(announceMd["recordingAnnouncement"]);
19674
- if (announcement) {
20114
+ if (announcement && callAdmission.kind !== "refused") {
19675
20115
  try {
19676
20116
  await adapter.say(announcement, { allowInterruptions: false });
19677
20117
  } catch (err) {
@@ -19679,7 +20119,13 @@ ${callIntent}`
19679
20119
  }
19680
20120
  }
19681
20121
  }
19682
- if (puppetOwnsOpening(this.config.puppetMode, flowBootFailure)) {
20122
+ if (callAdmission.kind === "refused") {
20123
+ console.warn("[agent] call refused at admission", { room: job.room.name ?? null, endReason: callAdmission.endReason });
20124
+ await playAdmissionRefusal(callAdmission, {
20125
+ say: (text, opts) => adapter.say(text, opts),
20126
+ hangUp: (reason) => endCallController.hangUp({ trigger: "admission", reason })
20127
+ });
20128
+ } else if (puppetOwnsOpening(this.config.puppetMode, flowBootFailure)) {
19683
20129
  const dispatchMd = parseJobMetadata(job.job?.metadata);
19684
20130
  const initialDirective = stringValue(dispatchMd["initialDirective"]);
19685
20131
  if (initialDirective) {
@@ -19771,6 +20217,7 @@ ${callIntent}`
19771
20217
  }
19772
20218
  await waitForSessionClose;
19773
20219
  await usageReported;
20220
+ await finalized;
19774
20221
  } catch (err) {
19775
20222
  console.error("[agent] runJob failed", {
19776
20223
  room: job.room.name ?? null,
@@ -19880,7 +20327,7 @@ function sharedBrainPubSub() {
19880
20327
  async function applyStoredPipeline(client, agentId, base, codeMode) {
19881
20328
  try {
19882
20329
  const stored = await client.agents.getConfig(agentId);
19883
- if (!stored) return { config: base, brain: "none", byokProviders: [] };
20330
+ if (!stored) return { config: base, brain: "none" };
19884
20331
  const mode = codeMode ?? stored.mode;
19885
20332
  const providers = /* @__PURE__ */ new Set();
19886
20333
  if (mode === "realtime" && stored.realtime) {
@@ -19955,7 +20402,6 @@ async function applyStoredPipeline(client, agentId, base, codeMode) {
19955
20402
  models: { ...base.models ?? {}, ...mapped, ...brainLlm ? { llm: brainLlm } : {} }
19956
20403
  },
19957
20404
  brain: brainLlm ? "bound" : brainConnectorId && mode !== "realtime" ? "unreachable" : "none",
19958
- byokProviders: Object.keys(creds2),
19959
20405
  ...mode !== "realtime" && stored.model ? { llm: { provider: stored.model.provider, model: stored.model.model } } : {}
19960
20406
  };
19961
20407
  } catch (err) {
@@ -19963,7 +20409,7 @@ async function applyStoredPipeline(client, agentId, base, codeMode) {
19963
20409
  agentId,
19964
20410
  err: err instanceof Error ? err.message : String(err)
19965
20411
  });
19966
- return { config: base, brain: "unknown", byokProviders: [] };
20412
+ return { config: base, brain: "unknown" };
19967
20413
  }
19968
20414
  }
19969
20415
  function firstString(...values) {