vent-hq 0.14.0 → 0.15.0

package/dist/index.mjs

@@ -94,7 +94,7 @@ function decodeJwt(jwt) {
 }
 
 // ../shared/src/constants.ts
-var RUNNER_CALLBACK_MAX_SKEW_MS = 5 * 6e4;
+var DEFAULT_AGENT_PORT = 3001;
 var CALLER_ACCENTS = [
   "american",
   "british",
@@ -4152,122 +4152,92 @@ var coerce = {
 };
 var NEVER = INVALID;
 
-// ../shared/src/schemas.ts
-var RefreshTokenRequestSchema = external_exports.object({
-  refresh_token: external_exports.string().regex(/^vent_[0-9a-f]{48}$/)
-});
-var AccessCredentialsSchema = external_exports.object({
-  access_token: external_exports.string().min(1),
-  refresh_token: RefreshTokenRequestSchema.shape.refresh_token
-});
-var CallerPersonaSchema = external_exports.object({
-  pace: external_exports.enum(["slow", "normal", "fast"]).optional(),
-  clarity: external_exports.enum(["clear", "vague", "rambling"]).optional(),
-  disfluencies: external_exports.boolean().optional(),
-  cooperation: external_exports.enum(["cooperative", "reluctant", "hostile"]).optional(),
-  emotion: external_exports.enum(["neutral", "cheerful", "confused", "frustrated", "skeptical", "rushed"]).optional(),
-  memory: external_exports.enum(["reliable", "unreliable"]).optional(),
-  intent_clarity: external_exports.enum(["clear", "indirect", "vague"]).optional(),
-  confirmation_style: external_exports.enum(["explicit", "vague"]).optional()
-}).optional();
-var CallerAudioEffectsSchema = external_exports.object({
-  noise: external_exports.object({
-    type: external_exports.enum(["babble", "white", "pink"]),
-    snr_db: external_exports.number().min(0).max(40).default(10)
-  }).optional(),
-  speed: external_exports.number().min(0.5).max(2).optional(),
-  speakerphone: external_exports.boolean().optional(),
-  mic_distance: external_exports.enum(["close", "normal", "far"]).optional(),
-  clarity: external_exports.number().min(0).max(1).optional(),
-  accent: external_exports.enum(CALLER_ACCENTS).optional(),
-  packet_loss: external_exports.number().min(0).max(0.3).optional(),
-  jitter_ms: external_exports.number().min(0).max(100).optional()
-});
-var ConversationCallSpecSchema = external_exports.object({
-  name: external_exports.string().optional(),
-  caller_prompt: external_exports.string().min(1),
-  max_turns: external_exports.number().int().min(1).max(50).default(6),
-  persona: CallerPersonaSchema,
-  caller_audio: CallerAudioEffectsSchema.optional(),
-  /** ISO 639-1 language code for multilingual calls (e.g., "es", "fr", "de"). Caller speaks this language, STT transcribes it, judge evaluates in it. */
-  language: external_exports.string().min(2).max(5).optional(),
-  /** Caller voice gender (English only; default female). Ignored if caller_audio.accent is set or language is non-English. */
-  voice: external_exports.enum(["male", "female"]).optional()
-});
-var CallSpecSchema = external_exports.object({
-  call: ConversationCallSpecSchema
-});
-var AdapterTypeSchema = external_exports.enum(["websocket", "livekit", "vapi", "retell", "elevenlabs", "bland"]);
-var PlatformProviderSchema = external_exports.enum(["bland", "livekit", "vapi", "retell", "elevenlabs"]);
-var ObservedToolCallSchema = external_exports.object({
-  name: external_exports.string(),
-  arguments: external_exports.record(external_exports.unknown()),
-  result: external_exports.unknown().optional(),
-  successful: external_exports.boolean().optional(),
-  provider_tool_type: external_exports.string().optional(),
-  timestamp_ms: external_exports.number().optional(),
-  latency_ms: external_exports.number().optional(),
-  turn_index: external_exports.number().int().min(0).optional()
-});
-var ToolCallMetricsSchema = external_exports.object({
-  total: external_exports.number().int().min(0),
-  successful: external_exports.number().int().min(0),
-  failed: external_exports.number().int().min(0),
-  mean_latency_ms: external_exports.number().optional(),
-  names: external_exports.array(external_exports.string())
-});
+// ../shared/src/schemas/platforms.ts
+var PLATFORM_PROVIDERS = ["vapi", "retell", "elevenlabs", "livekit", "bland"];
 var BasePlatformSchema = external_exports.object({
+  /** Max concurrent calls (platform-dependent) */
   max_concurrency: external_exports.number().int().min(1).optional()
 });
 var BlandPlatformSchema = BasePlatformSchema.extend({
   provider: external_exports.literal("bland"),
+  /** Bland API key (the CLI reads BLAND_API_KEY from the user's env) */
   bland_api_key: external_exports.string().optional(),
+  /** Bland pathway ID (the CLI reads BLAND_PATHWAY_ID from the user's env) */
   bland_pathway_id: external_exports.string().optional(),
+  /** Bland persona ID — reusable agent preset (the CLI reads BLAND_PERSONA_ID from the user's env) */
   persona_id: external_exports.string().optional(),
+  /** Task prompt — used instead of pathway_id for simple agents */
   task: external_exports.string().optional(),
+  /** Tool definitions (inline objects) or tool IDs (TL-xxx strings) */
   tools: external_exports.array(external_exports.unknown()).optional(),
+  /** Voice name ("maya", "josh") or custom voice clone UUID */
   voice: external_exports.string().optional(),
+  /** Model: "base" (full features), "turbo" (fastest, limited features) */
   model: external_exports.string().optional(),
+  /** Opening sentence — overrides any greeting in the task/pathway */
   first_sentence: external_exports.string().optional(),
+  /** If true, agent waits for callee to speak first (default: false) */
   wait_for_greeting: external_exports.boolean().optional(),
+  /** Max call duration in minutes (default: 30) */
   max_duration: external_exports.number().optional(),
+  /** Temperature 0-1 (default: 0.7) */
   temperature: external_exports.number().min(0).max(1).optional(),
+  /** Language code e.g. "babel-en", "babel-es" */
   language: external_exports.string().optional(),
+  /** How quickly agent stops speaking when interrupted, in ms (default: 500) */
   interruption_threshold: external_exports.number().optional(),
+  /** When true, agent ignores user interruptions entirely */
   block_interruptions: external_exports.boolean().optional(),
+  /** When true, enable Bland's noise filtering on caller audio */
   noise_cancellation: external_exports.boolean().optional(),
+  /** Background audio: "office", "cafe", "restaurant", "none", or null (default phone static) */
   background_track: external_exports.string().nullable().optional(),
+  /** Boost transcription accuracy for specific words. Supports "word:boost_factor" */
   keywords: external_exports.array(external_exports.string()).optional(),
+  /** Key-value pairs accessible as {{variable}} in agent prompts/pathways */
   request_data: external_exports.record(external_exports.unknown()).optional(),
+  /** Pronunciation overrides */
   pronunciation_guide: external_exports.array(external_exports.object({
     word: external_exports.string(),
     pronunciation: external_exports.string(),
     case_sensitive: external_exports.boolean().optional(),
     spaced: external_exports.boolean().optional()
   })).optional(),
+  /** Start pathway from a specific node instead of the default */
   start_node_id: external_exports.string().optional(),
+  /** Specific pathway version to test (default: production) */
   pathway_version: external_exports.number().optional()
 });
 var LiveKitPlatformSchema = BasePlatformSchema.extend({
   provider: external_exports.literal("livekit"),
+  /** LiveKit API key (the CLI reads LIVEKIT_API_KEY from the user's env) */
   livekit_api_key: external_exports.string().optional(),
+  /** LiveKit API secret (the CLI reads LIVEKIT_API_SECRET from the user's env) */
   livekit_api_secret: external_exports.string().optional(),
+  /** LiveKit server URL, e.g. wss://your-app.livekit.cloud (the CLI reads LIVEKIT_URL from the user's env) */
   livekit_url: external_exports.string().optional(),
+  /** Explicit agent dispatch — agent_name from WorkerOptions. Omit for automatic dispatch. */
   livekit_agent_name: external_exports.string().optional()
 }).strict();
 var VapiPlatformSchema = BasePlatformSchema.extend({
   provider: external_exports.literal("vapi"),
+  /** Vapi API key (the CLI reads VAPI_API_KEY from the user's env) */
   vapi_api_key: external_exports.string().optional(),
+  /** Vapi assistant ID (the CLI reads VAPI_ASSISTANT_ID or VAPI_AGENT_ID from the user's env) */
   vapi_assistant_id: external_exports.string().optional()
 });
 var RetellPlatformSchema = BasePlatformSchema.extend({
   provider: external_exports.literal("retell"),
+  /** Retell API key (the CLI reads RETELL_API_KEY from the user's env) */
   retell_api_key: external_exports.string().optional(),
+  /** Retell agent ID (the CLI reads RETELL_AGENT_ID from the user's env) */
   retell_agent_id: external_exports.string().optional()
 });
 var ElevenLabsPlatformSchema = BasePlatformSchema.extend({
   provider: external_exports.literal("elevenlabs"),
+  /** ElevenLabs API key (the CLI reads ELEVENLABS_API_KEY from the user's env) */
   elevenlabs_api_key: external_exports.string().optional(),
+  /** ElevenLabs agent ID (the CLI reads ELEVENLABS_AGENT_ID from the user's env) */
   elevenlabs_agent_id: external_exports.string().optional()
 });
 var PlatformConfigSchema = external_exports.discriminatedUnion("provider", [
@@ -4277,89 +4247,118 @@ var PlatformConfigSchema = external_exports.discriminatedUnion("provider", [
   RetellPlatformSchema,
   ElevenLabsPlatformSchema
 ]);
+var REQUIRED_PLATFORM_FIELDS = {
+  vapi: [["vapi_api_key"], ["vapi_assistant_id"]],
+  retell: [["retell_api_key"], ["retell_agent_id"]],
+  elevenlabs: [["elevenlabs_api_key"], ["elevenlabs_agent_id"]],
+  livekit: [["livekit_api_key"], ["livekit_api_secret"], ["livekit_url"]],
+  // Each test call creates a Bland web agent, which has no persona field: the
+  // instructions must come from a pathway or a task prompt, so a persona_id
+  // alone gives the agent no instructions.
+  bland: [["bland_api_key"], ["bland_pathway_id", "task"]]
+};
+function missingPlatformFields(platform) {
+  const values = platform;
+  const requirements = REQUIRED_PLATFORM_FIELDS[platform.provider];
+  return requirements.filter((anyOf) => anyOf.every((field) => !values[field]));
+}
+
+// ../shared/src/schemas/suite.ts
+var AdapterTypeSchema = external_exports.enum(["websocket", ...PLATFORM_PROVIDERS]);
+var CallerPersonaSchema = external_exports.object({
+  pace: external_exports.enum(["slow", "normal", "fast"]).optional(),
+  clarity: external_exports.enum(["clear", "vague", "rambling"]).optional(),
+  disfluencies: external_exports.boolean().optional(),
+  cooperation: external_exports.enum(["cooperative", "reluctant", "hostile"]).optional(),
+  emotion: external_exports.enum(["neutral", "cheerful", "confused", "frustrated", "skeptical", "rushed"]).optional(),
+  memory: external_exports.enum(["reliable", "unreliable"]).optional(),
+  intent_clarity: external_exports.enum(["clear", "indirect", "vague"]).optional(),
+  confirmation_style: external_exports.enum(["explicit", "vague"]).optional()
+});
+var CallerAudioEffectsSchema = external_exports.object({
+  noise: external_exports.object({
+    type: external_exports.enum(["babble", "white", "pink"]),
+    snr_db: external_exports.number().min(0).max(40).default(10)
+  }).optional(),
+  speed: external_exports.number().min(0.5).max(2).optional(),
+  speakerphone: external_exports.boolean().optional(),
+  mic_distance: external_exports.enum(["close", "normal", "far"]).optional(),
+  clarity: external_exports.number().min(0).max(1).optional(),
+  accent: external_exports.enum(CALLER_ACCENTS).optional(),
+  packet_loss: external_exports.number().min(0).max(0.3).optional(),
+  jitter_ms: external_exports.number().min(0).max(100).optional()
+});
+var ConversationCallSpecSchema = external_exports.object({
+  name: external_exports.string().optional(),
+  caller_prompt: external_exports.string().min(1),
+  max_turns: external_exports.number().int().min(1).max(50).default(6),
+  persona: CallerPersonaSchema.optional(),
+  /** Audio effects applied to all caller audio (speakerphone, speed, noise, accent, etc.) */
+  caller_audio: CallerAudioEffectsSchema.optional(),
+  /** ISO 639-1 language code for multilingual calls (e.g., "es", "fr", "de"). The caller speaks it and Vent transcribes in it. */
+  language: external_exports.string().min(2).max(5).optional(),
+  /** Caller voice gender (English only). "male" → aura-2-apollo-en, "female" → aura-2-thalia-en (default). Ignored if caller_audio.accent is set or language is non-English. */
+  voice: external_exports.enum(["male", "female"]).optional()
+});
+
+// ../shared/src/schemas/call-result.ts
+var ComponentLatencySchema = external_exports.object({
+  stt_ms: external_exports.number().optional(),
+  llm_ms: external_exports.number().optional(),
+  tts_ms: external_exports.number().optional(),
+  speech_duration_ms: external_exports.number().optional()
+});
 var ConversationTurnSchema = external_exports.object({
   role: external_exports.enum(["caller", "agent"]),
   text: external_exports.string(),
   timestamp_ms: external_exports.number(),
+  /** Caller decision mode that produced this turn, or the silent wait mode that led into an agent turn. */
   caller_decision_mode: external_exports.enum(["continue", "wait", "closing", "end_now"]).optional(),
   audio_duration_ms: external_exports.number().optional(),
+  /** Time to first byte: from the caller's audio being sent to the first agent audio arriving (ms). */
   ttfb_ms: external_exports.number().optional(),
+  /** Time to first word: from the caller's audio being sent to the agent starting to speak, by voice detection (ms). */
   ttfw_ms: external_exports.number().optional(),
+  /** Dead audio before speech starts: ttfw_ms - ttfb_ms */
   silence_pad_ms: external_exports.number().optional(),
   stt_confidence: external_exports.number().optional(),
+  /** Harness TTS synthesis time for this turn's caller audio (ms) */
   tts_ms: external_exports.number().optional(),
-  stt_ms: external_exports.number().optional(),
-  component_latency: external_exports.object({
-    stt_ms: external_exports.number().optional(),
-    llm_ms: external_exports.number().optional(),
-    tts_ms: external_exports.number().optional(),
-    speech_duration_ms: external_exports.number().optional()
-  }).optional(),
-  platform_transcript: external_exports.string().optional()
+  /** Harness STT transcription time for this turn's agent audio (ms) */
+  stt_ms: external_exports.number().optional()
 });
-var LatencyMetricsSchema = external_exports.object({
-  ttfb_per_turn_ms: external_exports.array(external_exports.number()),
-  p50_ttfb_ms: external_exports.number().optional(),
-  p90_ttfb_ms: external_exports.number().optional(),
-  p95_ttfb_ms: external_exports.number().optional(),
-  p99_ttfb_ms: external_exports.number().optional(),
-  first_turn_ttfb_ms: external_exports.number().optional(),
-  ttfw_per_turn_ms: external_exports.array(external_exports.number()).optional(),
-  p50_ttfw_ms: external_exports.number().optional(),
-  p90_ttfw_ms: external_exports.number().optional(),
-  p95_ttfw_ms: external_exports.number().optional(),
-  p99_ttfw_ms: external_exports.number().optional(),
-  first_turn_ttfw_ms: external_exports.number().optional(),
-  mean_silence_pad_ms: external_exports.number().optional(),
-  mouth_to_ear_est_ms: external_exports.number().optional(),
-  drift_slope_ms_per_turn: external_exports.number().optional()
+var PlatformTurnSchema = external_exports.object({
+  role: external_exports.enum(["caller", "agent"]),
+  /** Agent turns: the text the agent's LLM wrote. Caller turns: what the agent's STT heard the caller say. */
+  text: external_exports.string()
 });
-var HarnessOverheadSchema = external_exports.object({
-  tts_per_turn_ms: external_exports.array(external_exports.number()),
-  stt_per_turn_ms: external_exports.array(external_exports.number()),
-  mean_tts_ms: external_exports.number(),
-  mean_stt_ms: external_exports.number()
+var ObservedToolCallSchema = external_exports.object({
+  name: external_exports.string(),
+  arguments: external_exports.record(external_exports.unknown()),
+  result: external_exports.unknown().optional(),
+  successful: external_exports.boolean().optional(),
+  provider_tool_type: external_exports.string().optional(),
+  timestamp_ms: external_exports.number().optional(),
+  latency_ms: external_exports.number().optional(),
+  turn_index: external_exports.number().int().min(0).optional()
 });
 var SignalQualityMetricsSchema = external_exports.object({
+  /** Mean signal-to-noise ratio across turns (dB). >20 good, <10 bad */
   mean_snr_db: external_exports.number(),
+  /** Max clipping ratio across turns (0-1). >0.01 = distortion */
   max_clipping_ratio: external_exports.number(),
+  /** Mean energy consistency across turns (0-1). <0.5 = unstable volume */
   energy_consistency: external_exports.number(),
+  /** Total sudden volume drops across all turns */
   sudden_drops: external_exports.number().int().min(0),
+  /** Total sudden volume spikes across all turns */
   sudden_spikes: external_exports.number().int().min(0),
+  /** All turns have clean start/end (no clicks) */
   clean_edges: external_exports.boolean()
 });
-var ComponentLatencySchema = external_exports.object({
-  stt_ms: external_exports.number().optional(),
-  llm_ms: external_exports.number().optional(),
-  tts_ms: external_exports.number().optional(),
-  speech_duration_ms: external_exports.number().optional()
-});
-var ComponentLatencyMetricsSchema = external_exports.object({
-  per_turn: external_exports.array(ComponentLatencySchema),
-  mean_stt_ms: external_exports.number().optional(),
-  mean_llm_ms: external_exports.number().optional(),
-  mean_tts_ms: external_exports.number().optional(),
-  p95_stt_ms: external_exports.number().optional(),
-  p95_llm_ms: external_exports.number().optional(),
-  p95_tts_ms: external_exports.number().optional(),
-  bottleneck: external_exports.enum(["stt", "llm", "tts"]).optional()
-});
-var UsageEntrySchema = external_exports.object({
-  type: external_exports.string(),
-  provider: external_exports.string(),
-  model: external_exports.string(),
-  input_tokens: external_exports.number().int().optional(),
-  output_tokens: external_exports.number().int().optional()
-});
-var CostBreakdownSchema = external_exports.object({
-  stt_usd: external_exports.number().optional(),
-  llm_usd: external_exports.number().optional(),
-  tts_usd: external_exports.number().optional(),
-  transport_usd: external_exports.number().optional(),
-  platform_usd: external_exports.number().optional(),
-  total_usd: external_exports.number().optional(),
-  llm_prompt_tokens: external_exports.number().int().optional(),
-  llm_completion_tokens: external_exports.number().int().optional()
+var CallCostSchema = external_exports.object({
+  total_usd: external_exports.number(),
+  breakdown: external_exports.record(external_exports.number()).optional()
 });
 var ProviderWarningSchema = external_exports.object({
   message: external_exports.string().optional(),
@@ -4375,56 +4374,63 @@ var CallTransferSchema = external_exports.object({
 });
 var CallMetadataSchema = external_exports.object({
   platform: external_exports.string(),
+  /** Provider-visible call/conversation identifier for dashboard/API lookup. */
   provider_call_id: external_exports.string().optional(),
+  /** Secondary runtime/session identifier when distinct from the call ID. */
   provider_session_id: external_exports.string().optional(),
   ended_reason: external_exports.string().optional(),
-  cost_usd: external_exports.number().nullable().optional(),
-  cost_breakdown: CostBreakdownSchema.optional(),
+  cost: CallCostSchema.optional(),
   recording_url: external_exports.string().optional(),
+  /** Provider-specific alternate recording artifacts keyed by variant name. */
   recording_variants: external_exports.record(external_exports.string()).optional(),
+  /** Provider-specific debug and deep-link URLs keyed by purpose. */
   provider_debug_urls: external_exports.record(external_exports.string()).optional(),
-  variables: external_exports.record(external_exports.unknown()).optional(),
-  usage: external_exports.array(UsageEntrySchema).optional(),
+  /** Provider/runtime warnings retained in a normalized structure. */
   provider_warnings: external_exports.array(ProviderWarningSchema).optional(),
-  provider_metadata: external_exports.record(external_exports.unknown()).optional(),
+  /** Call transfers that occurred during the call */
   transfers: external_exports.array(CallTransferSchema).optional()
 });
 var ConversationMetricsSchema = external_exports.object({
-  mean_ttfb_ms: external_exports.number().optional(),
-  mean_ttfw_ms: external_exports.number().optional(),
-  latency: LatencyMetricsSchema.optional(),
-  tool_calls: ToolCallMetricsSchema.optional(),
+  /** How long Vent's connection to the platform took to open; feeds the mouth-to-ear estimate. */
+  connect_latency_ms: external_exports.number().optional(),
+  /** Raw audio signal quality (SNR, clipping, energy), aggregated across agent turns. */
   signal_quality: SignalQualityMetricsSchema.optional(),
-  harness_overhead: HarnessOverheadSchema.optional(),
-  component_latency: ComponentLatencyMetricsSchema.optional()
+  /** Every component timing the platform reported (STT/LLM/TTS). It can report
+   *  more than Vent has agent turns, so this is the full list. */
+  /** The platform's STT/LLM/TTS latencies for the call; the result shows their means. */
+  component_latency: external_exports.object({ per_turn: external_exports.array(ComponentLatencySchema) }).optional()
 });
 var ConversationCallResultSchema = external_exports.object({
   name: external_exports.string().optional(),
   caller_prompt: external_exports.string(),
   status: external_exports.enum(["completed", "error"]),
+  /** How a completed call ended, when Vent ended it or the connection closed. */
   end_note: external_exports.string().optional(),
+  /** Vent's record of the call: what Vent's caller said, and what Vent heard the agent say. */
   transcript: external_exports.array(ConversationTurnSchema),
+  /** The agent platform's own record of the call, when the platform provides one. */
+  platform_transcript: external_exports.array(PlatformTurnSchema).optional(),
   observed_tool_calls: external_exports.array(ObservedToolCallSchema).optional(),
   duration_ms: external_exports.number(),
   metrics: ConversationMetricsSchema,
   error: external_exports.string().optional(),
+  /** Platform call metadata (cost, ended reason, recording, analysis) */
   call_metadata: CallMetadataSchema.optional()
 });
-var RunAggregateV2Schema = external_exports.object({
-  conversation_calls: external_exports.object({
-    total: external_exports.number(),
-    passed: external_exports.number(),
-    failed: external_exports.number()
-  }).default({ total: 0, passed: 0, failed: 0 }),
-  total_duration_ms: external_exports.number(),
-  total_cost_usd: external_exports.number().optional()
+
+// ../shared/src/schemas/internal.ts
+var RefreshTokenRequestSchema = external_exports.object({
+  refresh_token: external_exports.string().regex(/^vent_[0-9a-f]{48}$/)
+});
+var AccessCredentialsSchema = external_exports.object({
+  access_token: external_exports.string().min(1),
+  refresh_token: RefreshTokenRequestSchema.shape.refresh_token
 });
-var RUNNER_TERMINAL_STATUSES = ["pass", "fail", "cancelled"];
-var RunnerCallbackV2Schema = external_exports.object({
+var RUN_TERMINAL_STATUSES = ["pass", "fail", "cancelled"];
+var WorkerResultCallbackSchema = external_exports.object({
   run_id: external_exports.string().uuid(),
-  status: external_exports.enum(RUNNER_TERMINAL_STATUSES),
+  status: external_exports.enum(RUN_TERMINAL_STATUSES),
   conversation_result: ConversationCallResultSchema,
-  aggregate: RunAggregateV2Schema,
   error_text: external_exports.string().optional()
 });
 var RunFailedCallbackSchema = external_exports.object({
@@ -4437,253 +4443,31 @@ var RunEventCallbackSchema = external_exports.object({
   message: external_exports.string().max(2e3),
   metadata_json: external_exports.record(external_exports.unknown()).optional()
 });
-var CallProgressCallbackSchema = external_exports.object({
+var RunJobSchema = external_exports.object({
   run_id: external_exports.string().uuid(),
-  completed: external_exports.number().int().nonnegative(),
-  total: external_exports.number().int().positive(),
-  call_type: external_exports.literal("conversation"),
-  call_name: external_exports.string().min(1).max(256),
-  status: external_exports.enum(["pass", "fail", "completed", "error"]),
-  duration_ms: external_exports.number().nonnegative(),
-  result: external_exports.record(external_exports.unknown()).optional()
+  user_id: external_exports.string().min(1),
+  adapter: AdapterTypeSchema,
+  call_spec: external_exports.object({ call: ConversationCallSpecSchema }),
+  agent_url: external_exports.string().optional(),
+  agent_session_id: external_exports.string().uuid().optional()
+});
+var AgentSessionResponseSchema = external_exports.object({
+  session_id: external_exports.string().uuid(),
+  relay_token: external_exports.string(),
+  api_url: external_exports.string(),
+  agent_port: external_exports.number().int().min(1).max(65535),
+  start_command: external_exports.string().nullable(),
+  health_endpoint: external_exports.string()
 });
-
-// ../shared/src/format-result.ts
-function formatConversationResult(raw, options = {}) {
-  if (!raw || typeof raw !== "object") return null;
-  const r = raw;
-  if (typeof r.caller_prompt !== "string") return null;
-  const verbose = options.verbose ?? false;
-  const debug = verbose ? formatDebug(r) : void 0;
-  const warnings = dedupeStrings([
-    ...formatProviderWarningMessages(r.call_metadata?.provider_warnings)
-  ]);
-  const result = {
-    name: r.name ?? null,
-    status: r.status,
-    duration_ms: r.duration_ms,
-    error: r.error ?? null,
-    end_note: r.end_note ?? null,
-    transcript: formatTranscript(r.transcript, options),
-    latency: r.metrics?.latency ? formatLatency(r.metrics.latency, r.metrics, verbose) : null,
-    tool_calls: formatToolCalls(r.metrics?.tool_calls, r.observed_tool_calls, verbose),
-    component_latency: formatComponentLatency(r.metrics?.component_latency, verbose),
-    call_metadata: formatCallMetadata(r.call_metadata, verbose)
-  };
-  if (warnings.length > 0) result.warnings = warnings;
-  if (verbose) result.caller_prompt = r.caller_prompt;
-  if (debug) result.debug = debug;
-  return result;
-}
-function formatTranscript(turns, options) {
-  if (!turns) return [];
-  return turns.map((t) => {
-    const turn = {
-      role: t.role,
-      text: t.text
-    };
-    if (t.ttfb_ms != null) turn.ttfb_ms = t.ttfb_ms;
-    if (t.ttfw_ms != null) turn.ttfw_ms = t.ttfw_ms;
-    if (t.audio_duration_ms != null) turn.audio_duration_ms = t.audio_duration_ms;
-    if (options.verbose) {
-      const debug = compactUnknownRecord({
-        timestamp_ms: t.timestamp_ms,
-        caller_decision_mode: t.caller_decision_mode,
-        silence_pad_ms: t.silence_pad_ms,
-        stt_confidence: t.stt_confidence,
-        harness_tts_ms: t.tts_ms,
-        harness_stt_ms: t.stt_ms,
-        component_latency: t.component_latency,
-        platform_transcript: t.platform_transcript
-      });
-      if (debug && Object.keys(debug).length > 0) {
-        turn.debug = debug;
-      }
-    }
-    return turn;
-  });
-}
-function formatLatency(latency, metrics, verbose) {
-  const hasTtfw = metrics.mean_ttfw_ms != null && latency.p50_ttfw_ms != null && latency.p95_ttfw_ms != null;
-  const responseTimeSource = hasTtfw ? "ttfw" : "ttfb";
-  const firstResponse = hasTtfw ? latency.first_turn_ttfw_ms ?? latency.first_turn_ttfb_ms : latency.first_turn_ttfb_ms;
-  if (firstResponse == null || !hasTtfw && (metrics.mean_ttfb_ms == null || latency.p95_ttfb_ms == null)) return null;
-  const result = {
-    response_time_ms: hasTtfw ? metrics.mean_ttfw_ms : metrics.mean_ttfb_ms,
-    response_time_source: responseTimeSource,
-    p95_response_time_ms: hasTtfw ? latency.p95_ttfw_ms : latency.p95_ttfb_ms,
-    first_response_time_ms: firstResponse
-  };
-  if (!verbose) return result;
-  result.p50_response_time_ms = hasTtfw ? latency.p50_ttfw_ms : latency.p50_ttfb_ms;
-  result.p90_response_time_ms = hasTtfw ? latency.p90_ttfw_ms ?? latency.p90_ttfb_ms : latency.p90_ttfb_ms;
-  result.p99_response_time_ms = hasTtfw ? latency.p99_ttfw_ms ?? latency.p99_ttfb_ms : latency.p99_ttfb_ms;
-  if (hasTtfw) {
-    result.mean_ttfw_ms = metrics.mean_ttfw_ms;
-    result.p50_ttfw_ms = latency.p50_ttfw_ms;
-    result.p90_ttfw_ms = latency.p90_ttfw_ms ?? latency.p90_ttfb_ms;
-    result.p95_ttfw_ms = latency.p95_ttfw_ms;
-    result.p99_ttfw_ms = latency.p99_ttfw_ms ?? latency.p99_ttfb_ms;
-    result.first_turn_ttfw_ms = latency.first_turn_ttfw_ms ?? latency.first_turn_ttfb_ms;
-  }
-  if (latency.drift_slope_ms_per_turn != null) result.drift_slope_ms_per_turn = latency.drift_slope_ms_per_turn;
-  if (latency.mean_silence_pad_ms != null) result.mean_silence_pad_ms = latency.mean_silence_pad_ms;
-  if (latency.mouth_to_ear_est_ms != null) result.mouth_to_ear_est_ms = latency.mouth_to_ear_est_ms;
-  return result;
-}
-function formatToolCalls(summary, observed, verbose) {
-  return {
-    total: summary?.total ?? observed?.length ?? 0,
-    successful: summary?.successful ?? observed?.filter((c) => c.successful).length ?? 0,
-    failed: summary?.failed ?? observed?.filter((c) => c.successful === false).length ?? 0,
-    mean_latency_ms: summary?.mean_latency_ms,
-    names: summary?.names ?? [...new Set((observed ?? []).map((c) => c.name))],
-    observed: (observed ?? []).map((c) => {
-      const entry = {
-        name: c.name,
-        arguments: stripExecutionMessage(c.arguments),
-        result: verbose ? c.result : truncateLargeArrays(c.result),
-        successful: c.successful,
-        latency_ms: c.latency_ms,
-        turn_index: c.turn_index
-      };
-      if (c.provider_tool_type && c.provider_tool_type !== "custom") {
-        entry.provider_tool_type = c.provider_tool_type;
-      }
-      return entry;
-    })
-  };
-}
-var TOOL_RESULT_ARRAY_LIMIT = 5;
-function truncateLargeArrays(value) {
-  if (Array.isArray(value)) {
-    if (value.length <= TOOL_RESULT_ARRAY_LIMIT) {
-      return value.map(truncateLargeArrays);
-    }
-    return [
-      ...value.slice(0, TOOL_RESULT_ARRAY_LIMIT).map(truncateLargeArrays),
-      { _truncated: `${value.length - TOOL_RESULT_ARRAY_LIMIT} more entries omitted; use --verbose for full result` }
-    ];
-  }
-  if (value && typeof value === "object") {
-    const out = {};
-    for (const [k, v] of Object.entries(value)) {
-      out[k] = truncateLargeArrays(v);
-    }
-    return out;
-  }
-  return value;
-}
-function stripExecutionMessage(args) {
-  if (!args || typeof args !== "object") return args;
-  const { execution_message: _drop, ...rest } = args;
-  return rest;
-}
-function formatComponentLatency(cl, verbose) {
-  if (!cl) return null;
-  const result = {
-    mean_stt_ms: cl.mean_stt_ms,
-    mean_llm_ms: cl.mean_llm_ms,
-    mean_tts_ms: cl.mean_tts_ms,
-    bottleneck: cl.bottleneck
-  };
-  if (!verbose) return result;
-  const speechDurations = cl.per_turn.map((t) => t.speech_duration_ms).filter((v) => v != null);
-  const meanSpeech = speechDurations.length > 0 ? Math.round(speechDurations.reduce((a, b) => a + b, 0) / speechDurations.length) : void 0;
-  result.p95_stt_ms = cl.p95_stt_ms;
-  result.p95_llm_ms = cl.p95_llm_ms;
-  result.p95_tts_ms = cl.p95_tts_ms;
-  result.mean_speech_duration_ms = meanSpeech;
-  return result;
-}
-function formatCallMetadata(meta, verbose) {
-  if (!meta) return null;
-  const transfers = meta.transfers?.map((transfer) => {
-    const formattedTransfer = {
-      type: transfer.type,
-      destination: transfer.destination,
-      status: transfer.status,
-      sources: transfer.sources
-    };
-    if (transfer.timestamp_ms != null) {
-      formattedTransfer.timestamp_ms = transfer.timestamp_ms;
-    }
-    return formattedTransfer;
-  });
-  const result = {
-    platform: meta.platform,
-    provider_call_id: meta.provider_call_id,
-    provider_session_id: meta.provider_session_id,
-    ended_reason: meta.ended_reason,
-    cost_usd: meta.cost_usd,
-    recording_url: meta.recording_url
-  };
-  if (verbose) {
-    result.cost_breakdown = meta.cost_breakdown;
-    result.usage = meta.usage && meta.usage.length > 0 ? meta.usage : void 0;
-    result.recording_variants = meta.recording_variants;
-    result.provider_debug_urls = meta.provider_debug_urls;
-    result.variables = meta.variables;
-  }
-  if (transfers && transfers.length > 0) {
-    const completedTransferCount = transfers.filter((transfer) => transfer.status === "completed").length;
-    const transferCompleted = completedTransferCount > 0;
-    result.transfer_attempted = true;
-    result.transfer_completed = transferCompleted;
-    result.escalated = transferCompleted;
-    result.transfer_count = transfers.length;
-    result.completed_transfer_count = completedTransferCount;
-    result.transfers = transfers;
-  }
-  return result;
-}
-function formatDebug(result) {
-  const debug = compactUnknownRecord({
-    signal_quality: result.metrics?.signal_quality,
-    harness_overhead: result.metrics?.harness_overhead,
-    provider_warnings: nonEmptyArray(result.call_metadata?.provider_warnings),
-    component_latency_per_turn: nonEmptyArray(result.metrics?.component_latency?.per_turn),
-    observed_tool_calls: formatDebugToolCalls(result.observed_tool_calls),
-    provider_metadata: result.call_metadata?.provider_metadata
-  });
-  return debug && Object.keys(debug).length > 0 ? debug : void 0;
-}
-function formatDebugToolCalls(observed) {
-  if (!observed || observed.length === 0) return void 0;
-  return observed.map((call) => ({
-    name: call.name,
-    arguments: call.arguments,
-    result: call.result,
-    successful: call.successful,
-    provider_tool_type: call.provider_tool_type,
-    timestamp_ms: call.timestamp_ms,
-    latency_ms: call.latency_ms,
-    turn_index: call.turn_index
-  }));
-}
-function nonEmptyArray(value) {
-  return value && value.length > 0 ? value : void 0;
-}
-function formatProviderWarningMessages(warnings) {
-  if (!warnings || warnings.length === 0) return [];
-  return warnings.map((warning) => warning.message ?? warning.code).filter((message) => typeof message === "string" && message.length > 0);
-}
-function dedupeStrings(values) {
-  return [...new Set(values)];
-}
-function compactUnknownRecord(record) {
-  const entries = Object.entries(record).filter(([, value]) => value != null);
-  return entries.length > 0 ? Object.fromEntries(entries) : void 0;
-}
 
 // ../shared/src/callback-auth.ts
 import { createHmac, timingSafeEqual } from "node:crypto";
+var RUNNER_CALLBACK_MAX_SKEW_MS = 5 * 6e4;
 
 // src/lib/config.ts
 var CONFIG_DIR = path.join(homedir(), ".vent");
 var CREDENTIALS_FILE = path.join(CONFIG_DIR, "credentials");
 var API_BASE = process.env.VENT_API_URL ?? "https://api.venthq.dev";
-var DASHBOARD_URL = process.env.VENT_DASHBOARD_URL ?? "https://venthq.dev";
 var refreshes = /* @__PURE__ */ new Map();
 async function readCredentials() {
   const accessToken = process.env.VENT_ACCESS_TOKEN;
@@ -4751,6 +4535,11 @@ async function loadAccessToken() {
   if (source.refreshToken && !usableAccessToken(source.accessToken, 6e4)) return refresh(source);
   return usableAccessToken(source.accessToken) ? source.accessToken : null;
 }
+async function requireAccessToken() {
+  const token = await loadAccessToken();
+  if (!token) throw new Error("No Vent credentials found. Set VENT_REFRESH_TOKEN or run `npx vent-hq login`.");
+  return token;
+}
 async function refreshAccessToken(rejectedToken) {
   const source = await readCredentials();
   if (!source?.refreshToken) return null;
@@ -4785,9 +4574,9 @@ var ApiError = class extends Error {
     this.body = body;
   }
 };
-async function authenticatedFetch(path4, apiKey, options = {}) {
+async function authenticatedFetch(path4, options = {}) {
   const url = `${API_BASE}${path4}`;
-  const token = await loadAccessToken() ?? apiKey;
+  const token = await requireAccessToken();
   const send = (accessToken) => {
     const headers = new Headers(options.headers);
     headers.set("Authorization", `Bearer ${accessToken}`);
@@ -4801,8 +4590,8 @@ async function authenticatedFetch(path4, apiKey, options = {}) {
   await response.body?.cancel();
   return send(refreshed);
 }
-async function apiFetch(path4, apiKey, options = {}) {
-  const res = await authenticatedFetch(path4, apiKey, options);
+async function apiFetch(path4, options = {}) {
+  const res = await authenticatedFetch(path4, options);
   if (!res.ok) {
     const text = await res.text();
     let body = text;
@@ -4818,15 +4607,15 @@ async function apiFetch(path4, apiKey, options = {}) {
 // src/lib/submit-run.ts
 import { randomUUID as randomUUID2 } from "node:crypto";
 var RETRY_DELAY_MS = 2e3;
-async function submitRun(accessToken, payload) {
+async function submitRun(payload) {
   const body = JSON.stringify({ ...payload, idempotency_key: randomUUID2() });
-  const send = () => apiFetch("/runs/submit", accessToken, { method: "POST", body });
+  const send = () => apiFetch("/runs/submit", { method: "POST", body });
   try {
     return await send();
   } catch (err) {
     const answerMayBeLost = !(err instanceof ApiError) || [502, 503, 504].includes(err.status);
     if (!answerMayBeLost) throw err;
-    await new Promise((resolve2) => setTimeout(resolve2, RETRY_DELAY_MS));
+    await new Promise((resolve) => setTimeout(resolve, RETRY_DELAY_MS));
     return send();
   }
 }
@@ -4834,7 +4623,7 @@ async function submitRun(accessToken, payload) {
 // src/lib/sse.ts
 var MAX_RETRIES = 5;
 var RETRY_DELAY_MS2 = 2e3;
-async function* streamRunEvents(runId, apiKey, signal) {
+async function* streamRunEvents(runId, signal) {
   const path4 = `/runs/${runId}/stream`;
   const seenIds = /* @__PURE__ */ new Set();
   let retries = 0;
@@ -4844,9 +4633,7 @@ async function* streamRunEvents(runId, apiKey, signal) {
     }
     let res;
     try {
-      res = await authenticatedFetch(path4, apiKey, {
-        signal
-      });
+      res = await authenticatedFetch(path4, { signal });
     } catch (err) {
       if (err.name === "AbortError") throw err;
       retries++;
@@ -4933,29 +4720,13 @@ var red = (s) => isTTY ? `\x1B[31m${s}\x1B[0m` : s;
 var yellow = (s) => isTTY ? `\x1B[33m${s}\x1B[0m` : s;
 var blue = (s) => isTTY ? `\x1B[34m${s}\x1B[0m` : s;
 function printEvent(event) {
-  if (!isTTY) return;
-  const meta = event.metadata_json ?? {};
-  switch (event.event_type) {
-    case "call_completed":
-      printCallResult(meta);
-      break;
-    case "run_complete":
-      printRunComplete(meta);
-      break;
-    case "call_started": {
-      const name = meta.call_name ?? "call";
-      stdoutSync(dim(`  \u25B8 ${name}\u2026`) + "\n");
-      break;
-    }
-    default:
-      stdoutSync(dim(`  [${event.event_type}]`) + "\n");
-  }
+  if (!isTTY || event.event_type === "run_complete") return;
+  stdoutSync(dim(`  [${event.event_type}]`) + "\n");
 }
-function printCallResult(meta) {
-  const result = meta.result;
-  const callName = result?.name ?? meta.call_name ?? "call";
-  const callStatus = result?.status ?? meta.status;
-  const durationMs = result?.duration_ms ?? meta.duration_ms;
+function printCallResult(result) {
+  const callName = result.name ?? "call";
+  const callStatus = result.status;
+  const durationMs = result.duration_ms;
   const statusIcon = callStatus === "completed" ? green("\u25CF") : red("\u25CF");
   const duration = durationMs != null ? (durationMs / 1e3).toFixed(1) + "s" : "\u2014";
   const parts = [statusIcon, bold(callName), dim(duration)];
@@ -4985,113 +4756,21 @@ function printCallResult(meta) {
     }
   }
 }
-function printRunComplete(meta) {
-  const agg = meta.aggregate;
-  const counts = agg?.conversation_calls;
-  const total = meta.total_calls ?? counts?.total;
-  stdoutSync("\n");
-  stdoutSync(bold("Run complete") + dim(" \u2014 Vent does not judge mission success; review the calls above.") + "\n");
-  if (total != null) {
-    stdoutSync(`${total} call${total === 1 ? "" : "s"} ran
-`);
-  }
-  if (typeof meta.error_text === "string") {
-    stdoutSync(red(`Run ${String(meta.status ?? "ended")}: ${meta.error_text}`) + "\n");
-  }
-}
-function printSummary(callResults, runComplete, runId, options = {}) {
-  const allCalls = options.rawCalls ? formatRawCalls(options.rawCalls, options.verbose ?? false) : callResults.map((e) => {
-    const meta = e.metadata_json ?? {};
-    const r = meta.result;
-    if (r) return r;
-    return {
-      name: meta.call_name ?? "call",
-      status: meta.status ?? "unknown",
-      duration_ms: meta.duration_ms,
-      error: null
-    };
-  });
-  const agg = runComplete.aggregate;
-  const counts = agg?.conversation_calls;
-  const summaryData = buildRunSummaryJson({
-    runId,
-    status: runComplete.status,
-    total: runComplete.total_calls ?? counts?.total,
-    formattedCalls: allCalls,
-    verbose: options.verbose,
-    runDetails: options.runDetails ?? {
-      aggregate: runComplete.aggregate,
-      error_text: runComplete.error_text
-    }
-  });
+function printRunOutput(output) {
   if (!isTTY) {
-    stdoutSync(JSON.stringify(summaryData, null, 2) + "\n");
+    stdoutSync(JSON.stringify(output, null, 2) + "\n");
     return;
   }
-  const errored = allCalls.filter((t) => t.status && t.status !== "completed");
-  if (errored.length > 0) {
-    stdoutSync("\n" + bold("Pipeline errors:") + "\n");
-    for (const t of errored) {
-      const ms = typeof t.duration_ms === "number" ? t.duration_ms : null;
-      const duration = ms != null ? (ms / 1e3).toFixed(1) + "s" : "\u2014";
-      const name = typeof t.name === "string" ? t.name : "call";
-      const parts = [red("\u25CF"), bold(name), dim(duration)];
-      stdoutSync("  " + parts.join("  ") + "\n");
-    }
-  }
-  stdoutSync(dim(`Run ID: ${runId}`) + "\n");
-}
-function buildRunSummaryJson(options) {
-  const calls = options.rawCalls ? formatRawCalls(options.rawCalls, options.verbose ?? false) : options.formattedCalls ?? [];
-  const summaryData = {
-    run_id: options.runId,
-    ...options.status != null ? { status: options.status } : {},
-    total: options.total,
-    ...options.passed != null ? { passed: options.passed } : {},
-    ...options.failed != null ? { failed: options.failed } : {},
-    calls
-  };
-  const details = options.runDetails;
-  if (details?.created_at != null) summaryData["created_at"] = details.created_at;
-  if (details?.started_at != null) summaryData["started_at"] = details.started_at;
-  if (details?.finished_at != null) summaryData["finished_at"] = details.finished_at;
-  if (details?.duration_ms != null) summaryData["duration_ms"] = details.duration_ms;
-  if (details?.error_text != null) summaryData["error_text"] = details.error_text;
-  if (details?.aggregate != null) {
-    summaryData["aggregate"] = stripPassFailFromAggregate(details.aggregate);
-  }
-  return summaryData;
-}
-function stripPassFailFromAggregate(aggregate) {
-  if (!aggregate || typeof aggregate !== "object" || Array.isArray(aggregate)) return aggregate;
-  const out = {};
-  for (const [k, v] of Object.entries(aggregate)) {
-    if (k === "conversation_calls" && v && typeof v === "object") {
-      const cc = v;
-      const filtered = {};
-      for (const [ck, cv] of Object.entries(cc)) {
-        if (ck === "passed" || ck === "failed") continue;
-        filtered[ck] = cv;
-      }
-      out[k] = filtered;
-    } else {
-      out[k] = v;
-    }
+  const calls = output.calls ?? [];
+  for (const call of calls) printCallResult(call);
+  stdoutSync("\n");
+  stdoutSync(bold("Run complete") + dim(" \u2014 Vent does not judge mission success; review the calls above.") + "\n");
+  stdoutSync(`${calls.length} call${calls.length === 1 ? "" : "s"} ran
+`);
+  if (typeof output.error_text === "string") {
+    stdoutSync(red(`Run ${String(output.status ?? "ended")}: ${output.error_text}`) + "\n");
   }
-  return out;
-}
-function formatRawCalls(rawCalls, verbose) {
-  return rawCalls.map((raw) => {
-    const formatted = formatConversationResult(raw, { verbose });
-    if (formatted) return formatted;
-    const fallback = raw;
-    return {
-      name: typeof fallback["name"] === "string" ? fallback["name"] : "call",
-      status: typeof fallback["status"] === "string" ? fallback["status"] : "unknown",
-      duration_ms: typeof fallback["duration_ms"] === "number" ? fallback["duration_ms"] : void 0,
-      error: typeof fallback["error"] === "string" ? fallback["error"] : null
-    };
-  });
+  stdoutSync(dim(`Run ID: ${String(output.run_id)}`) + "\n");
 }
 function printError(message) {
   const line = red(bold("error")) + ` ${message}
@@ -5128,35 +4807,25 @@ function gitSha() {
     return null;
   }
 }
-async function saveRunHistory(runId, callResults, runCompleteData) {
+async function saveRunHistory(output) {
   try {
+    const runId = output.run_id;
     const dir = path2.join(process.cwd(), ".vent", "runs");
     await fs2.mkdir(dir, { recursive: true });
     const now = /* @__PURE__ */ new Date();
     const timestamp = now.toISOString().replace(/[:.]/g, "-").slice(0, 19);
     const shortId = runId.slice(0, 8);
-    const aggregate = runCompleteData.aggregate;
-    const convCalls = aggregate?.conversation_calls;
-    const total = convCalls?.total ?? 0;
-    const passed = convCalls?.passed ?? 0;
-    const failed = convCalls?.failed ?? 0;
+    const calls = output.calls;
     const entry = {
       run_id: runId,
       timestamp: now.toISOString(),
       git_sha: gitSha(),
       summary: {
-        status: runCompleteData.status ?? "unknown",
-        calls_total: total,
-        calls_passed: passed,
-        calls_failed: failed,
-        total_duration_ms: aggregate?.total_duration_ms,
-        total_cost_usd: aggregate?.total_cost_usd
+        status: output.status,
+        calls_total: calls.length,
+        total_duration_ms: calls.reduce((sum, c) => sum + (c.duration_ms ?? 0), 0)
       },
-      call_results: callResults.map((e) => {
-        const meta = e.metadata_json ?? {};
-        const result = meta.result ?? {};
-        return { ...result, status: meta.status };
-      })
+      call_results: calls
     };
     const filename = `${timestamp}_${shortId}.json`;
     const filepath = path2.join(dir, filename);
@@ -5181,55 +4850,16 @@ function isRemotePlatformAdapter(adapter) {
 function looksLikeEnvReference(value) {
   return typeof value === "string" && /^[A-Z][A-Z0-9_]+$/.test(value);
 }
-function isMissingResolvedValue(value) {
-  return !value || looksLikeEnvReference(value);
+function describeField(adapter, field) {
+  if (adapter === "vapi" && field === "vapi_assistant_id") return "VAPI_ASSISTANT_ID or VAPI_AGENT_ID";
+  const envFields = PLATFORM_ENV_MAP[adapter];
+  return envFields[field] ?? `platform.${field}`;
 }
-function requireField(platform) {
-  switch (platform.provider) {
-    case "vapi":
-      if (isMissingResolvedValue(platform.vapi_api_key)) {
-        throw new Error("Missing VAPI_API_KEY in .env");
-      }
-      if (isMissingResolvedValue(platform.vapi_assistant_id)) {
-        throw new Error("Missing VAPI_ASSISTANT_ID or VAPI_AGENT_ID in .env");
-      }
-      return;
-    case "retell":
-      if (isMissingResolvedValue(platform.retell_api_key)) {
-        throw new Error("Missing RETELL_API_KEY in .env");
-      }
-      if (isMissingResolvedValue(platform.retell_agent_id)) {
-        throw new Error("Missing RETELL_AGENT_ID in .env");
-      }
-      return;
-    case "elevenlabs":
-      if (isMissingResolvedValue(platform.elevenlabs_api_key)) {
-        throw new Error("Missing ELEVENLABS_API_KEY in .env");
-      }
-      if (isMissingResolvedValue(platform.elevenlabs_agent_id)) {
-        throw new Error("Missing ELEVENLABS_AGENT_ID in .env");
-      }
-      return;
-    case "livekit":
-      if (isMissingResolvedValue(platform.livekit_api_key)) {
-        throw new Error("Missing LIVEKIT_API_KEY in .env");
-      }
-      if (isMissingResolvedValue(platform.livekit_api_secret)) {
-        throw new Error("Missing LIVEKIT_API_SECRET in .env");
-      }
-      if (isMissingResolvedValue(platform.livekit_url)) {
-        throw new Error("Missing LIVEKIT_URL in .env");
-      }
-      return;
-    case "bland":
-      if (isMissingResolvedValue(platform.bland_api_key)) {
-        throw new Error("Missing BLAND_API_KEY in .env");
-      }
-      if (isMissingResolvedValue(platform.bland_pathway_id) && isMissingResolvedValue(platform.task) && isMissingResolvedValue(platform.persona_id)) {
-        throw new Error("Missing BLAND_PATHWAY_ID, BLAND_PERSONA_ID, or platform.task in .env");
-      }
-      return;
-  }
+function requireFields(adapter, platform) {
+  const missing = missingPlatformFields(platform);
+  if (missing.length === 0) return;
+  const list = missing.map((anyOf) => anyOf.map((field) => describeField(adapter, field)).join(" or ")).join(", ");
+  throw new Error(`Missing ${list} (set it in .env, .env.local or the environment)`);
 }
 function resolveRemotePlatformConfig(config) {
   const cfg = config;
@@ -5247,12 +4877,8 @@ function resolveRemotePlatformConfig(config) {
   const envFields = PLATFORM_ENV_MAP[adapter];
   for (const [field, envVar] of Object.entries(envFields)) {
     const currentValue = resolved[field];
-    const needsResolve = !currentValue || currentValue === envVar || looksLikeEnvReference(currentValue);
-    if (needsResolve) {
-      const envValue = process.env[envVar];
-      if (envValue) {
-        resolved[field] = envValue;
-      }
+    if (!currentValue || looksLikeEnvReference(currentValue)) {
+      resolved[field] = process.env[envVar] || void 0;
     }
   }
   if (adapter === "vapi" && !resolved.vapi_assistant_id) {
@@ -5260,7 +4886,7 @@ function resolveRemotePlatformConfig(config) {
     if (alt) resolved.vapi_assistant_id = alt;
   }
   const platform = PlatformConfigSchema.parse(resolved);
-  requireField(platform);
+  requireFields(adapter, platform);
   return platform;
 }
 
@@ -5320,15 +4946,10 @@ async function runCommand(args) {
     );
     return 2;
   }
-  const accessToken = await loadAccessToken();
-  if (!accessToken) {
-    printError("No Vent credentials found. Set VENT_REFRESH_TOKEN or run `npx vent-hq login`.");
-    return 2;
-  }
   printInfo("Submitting run\u2026");
   let submitResult;
   try {
-    const res = await submitRun(accessToken, { config, agent_session_id: args.session });
+    const res = await submitRun({ config, agent_session_id: args.session });
     submitResult = await res.json();
   } catch (err) {
     if (err instanceof ApiError && err.status === 401) {
@@ -5347,14 +4968,13 @@ async function runCommand(args) {
   printInfo(`Streaming results for run ${run_id}\u2026`);
   const abortController = new AbortController();
   let exitCode = 0;
-  const callResults = [];
-  let runCompleteData = null;
+  let runOutput = null;
   let stopRequested = false;
   let stopPromise = null;
   const onSignal = () => {
     if (!stopRequested) {
       stopRequested = true;
-      stopPromise = apiFetch(`/runs/${run_id}/stop`, accessToken, {
+      stopPromise = apiFetch(`/runs/${run_id}/stop`, {
         method: "POST",
         signal: AbortSignal.timeout(5e3)
       }).catch(() => {
@@ -5365,15 +4985,10 @@ async function runCommand(args) {
   process.on("SIGINT", onSignal);
   process.on("SIGTERM", onSignal);
   try {
-    for await (const event of streamRunEvents(run_id, accessToken, abortController.signal)) {
-      const meta = event.metadata_json ?? {};
+    for await (const event of streamRunEvents(run_id, abortController.signal)) {
       printEvent(event);
-      if (event.event_type === "call_completed") {
-        callResults.push(event);
-      }
       if (event.event_type === "run_complete") {
-        runCompleteData = meta;
-        exitCode = meta["status"] === "pass" ? 0 : 1;
+        runOutput = event.result ?? null;
       }
       if (event.event_type === "error") {
         printError(event.message ?? "Stream connection lost");
@@ -5389,34 +5004,22 @@ async function runCommand(args) {
     process.off("SIGINT", onSignal);
     process.off("SIGTERM", onSignal);
   }
-  if (runCompleteData) {
-    let rawRunDetails = null;
-    if (args.verbose && !isTTY2 || callResults.length === 0) {
-      try {
-        const res = await apiFetch(`/runs/${run_id}`, accessToken);
-        rawRunDetails = await res.json();
-      } catch {
-        printWarn("Run result fetch failed; showing the streamed summary only.");
-      }
+  if (!runOutput && !stopRequested) {
+    try {
+      const res = await apiFetch(`/runs/${run_id}`);
+      const stored = await res.json();
+      if (["pass", "fail", "cancelled"].includes(String(stored.status))) runOutput = stored;
+    } catch {
+      printWarn("Run result fetch failed.");
     }
-    printSummary(callResults, runCompleteData, run_id, {
-      verbose: args.verbose,
-      rawCalls: Array.isArray(rawRunDetails?.["results"]) ? rawRunDetails["results"] : void 0,
-      runDetails: rawRunDetails ? {
-        created_at: rawRunDetails["created_at"],
-        started_at: rawRunDetails["started_at"],
-        finished_at: rawRunDetails["finished_at"],
-        duration_ms: rawRunDetails["duration_ms"],
-        error_text: rawRunDetails["error_text"],
-        aggregate: rawRunDetails["aggregate"]
-      } : void 0
-    });
   }
-  if (runCompleteData) {
-    const savedPath = await saveRunHistory(run_id, callResults, runCompleteData);
-    if (savedPath) {
-      printInfo(`Run saved to ${savedPath}`);
-    }
+  if (runOutput) {
+    exitCode = runOutput.status === "pass" ? 0 : 1;
+    printRunOutput(runOutput);
+    const savedPath = await saveRunHistory(runOutput);
+    if (savedPath) printInfo(`Run saved to ${savedPath}`);
+  } else if (exitCode === 0) {
+    exitCode = 2;
   }
   if (stopPromise) {
     await stopPromise;
@@ -5431,7 +5034,11 @@ import * as net from "node:net";
 // src/lib/relay.ts
 import { spawn } from "node:child_process";
 
-// ../relay-client/src/client.ts
+// src/lib/relay-client.ts
+var CONNECT_TIMEOUT_MS = 3e4;
+function errorEventMessage(ev) {
+  return ev.message;
+}
 var RelayClient = class {
   controlWs = null;
   localConnections = /* @__PURE__ */ new Map();
@@ -5451,10 +5058,10 @@ var RelayClient = class {
       handler(...args);
     }
   }
-  async connect(timeoutMs = 3e4) {
+  async connect() {
     const wsBase = this.config.apiUrl.replace(/^http/, "ws");
     const controlUrl = `${wsBase}/relay/control?session_id=${this.config.sessionId}&token=${this.config.relayToken}`;
-    return new Promise((resolve2, reject) => {
+    return new Promise((resolve, reject) => {
       const ws = new WebSocket(controlUrl);
       ws.binaryType = "arraybuffer";
       let settled = false;
@@ -5468,29 +5075,28 @@ var RelayClient = class {
       const hardTimeout = setTimeout(() => {
         settle(() => {
           ws.close();
-          reject(new Error(`Relay connection timed out after ${timeoutMs}ms \u2014 check network connectivity to ${this.config.apiUrl}`));
+          reject(new Error(`Relay connection timed out after ${CONNECT_TIMEOUT_MS}ms \u2014 check network connectivity to ${this.config.apiUrl}`));
         });
-      }, timeoutMs);
+      }, CONNECT_TIMEOUT_MS);
       ws.addEventListener("open", () => {
         this.controlWs = ws;
         this.setupControlHandlers(ws);
-        this.emit("connected");
       });
-      this.on("ready", () => settle(() => resolve2()));
+      this.on("ready", () => settle(() => resolve()));
       ws.addEventListener("close", (ev) => {
         settle(() => reject(new Error(`Relay rejected the connection (${ev.code}${ev.reason ? `: ${ev.reason}` : ""})`)));
       });
       ws.addEventListener("error", (ev) => {
         if (!this.controlWs) {
-          settle(() => reject(new Error(`Failed to connect to relay: ${ev.message ?? "connection error"}`)));
+          settle(() => reject(new Error(`Failed to connect to relay: ${errorEventMessage(ev) ?? "connection error"}`)));
         }
       });
     });
   }
   async disconnect() {
     this.closed = true;
-    for (const [connId, conn] of this.localConnections) {
-      if (conn.local.readyState !== WebSocket.CLOSED) conn.local.close();
+    for (const [connId, local] of this.localConnections) {
+      if (local.readyState !== WebSocket.CLOSED) local.close();
       this.localConnections.delete(connId);
     }
     if (this.controlWs) {
@@ -5523,12 +5129,12 @@ var RelayClient = class {
         if (frameType !== 1 && frameType !== 2) return;
         const connId = new TextDecoder().decode(data.subarray(1, 37));
         const payload = data.subarray(37);
-        const conn = this.localConnections.get(connId);
-        if (conn?.local.readyState === WebSocket.OPEN) {
+        const local = this.localConnections.get(connId);
+        if (local?.readyState === WebSocket.OPEN) {
           if (frameType === 2) {
-            conn.local.send(new TextDecoder().decode(payload));
+            local.send(new TextDecoder().decode(payload));
           } else {
-            conn.local.send(payload);
+            local.send(payload);
           }
         }
         return;
@@ -5541,8 +5147,8 @@ var RelayClient = class {
         } else if (msg.type === "new_connection" && msg.conn_id) {
           this.handleNewConnection(msg.conn_id);
         } else if (msg.type === "close" && msg.conn_id) {
-          const conn = this.localConnections.get(msg.conn_id);
-          if (conn?.local.readyState !== WebSocket.CLOSED) conn?.local.close();
+          const local = this.localConnections.get(msg.conn_id);
+          if (local && local.readyState !== WebSocket.CLOSED) local.close();
           this.localConnections.delete(msg.conn_id);
         }
       } catch {
@@ -5559,58 +5165,44 @@ var RelayClient = class {
       }
     });
     ws.addEventListener("error", (ev) => {
-      this.emit("error", new Error(ev.message ?? "WebSocket error"));
+      this.emit("error", new Error(errorEventMessage(ev) ?? "WebSocket error"));
     });
   }
   handleNewConnection(connId) {
-    const agentUrl = `ws://localhost:${this.config.agentPort}`;
-    this.emit("log", `[relay] new_connection ${connId} \u2192 connecting to ${agentUrl}`);
-    try {
-      const localWs = new WebSocket(agentUrl);
-      localWs.binaryType = "arraybuffer";
-      localWs.addEventListener("open", () => {
-        this.emit("log", `[relay] local WS open for ${connId}`);
-        this.sendControlMessage({ type: "open_ack", conn_id: connId });
-        this.localConnections.set(connId, { local: localWs, connId });
-      });
-      localWs.addEventListener("message", (event) => {
-        if (event.data instanceof ArrayBuffer) {
-          this.sendDataFrame(connId, new Uint8Array(event.data), 1);
-        } else {
-          this.sendDataFrame(connId, new TextEncoder().encode(event.data), 2);
-        }
-      });
-      let cleaned = false;
-      const cleanup = (reason) => {
-        if (cleaned) return;
-        cleaned = true;
-        this.emit("log", `[relay] local WS cleanup for ${connId}: ${reason ?? "unknown"}`);
-        if (localWs.readyState !== WebSocket.CLOSED) localWs.close();
-        this.localConnections.delete(connId);
-        this.sendControlMessage({ type: "close", conn_id: connId });
-      };
-      localWs.addEventListener("close", () => cleanup("close"));
-      localWs.addEventListener("error", (ev) => {
-        const msg = ev.message ?? "unknown error";
-        this.emit("log", `[relay] local WS error for ${connId}: ${msg}`);
-        cleanup(`error: ${msg}`);
-      });
-    } catch (err) {
-      this.emit("log", `[relay] Failed to connect local agent for ${connId}: ${err}`);
-    }
+    const localWs = new WebSocket(`ws://localhost:${this.config.agentPort}`);
+    localWs.binaryType = "arraybuffer";
+    localWs.addEventListener("open", () => {
+      this.sendControlMessage({ type: "open_ack", conn_id: connId });
+      this.localConnections.set(connId, localWs);
+    });
+    localWs.addEventListener("message", (event) => {
+      if (event.data instanceof ArrayBuffer) {
+        this.sendDataFrame(connId, new Uint8Array(event.data), 1);
+      } else {
+        this.sendDataFrame(connId, new TextEncoder().encode(event.data), 2);
+      }
+    });
+    let cleaned = false;
+    const cleanup = () => {
+      if (cleaned) return;
+      cleaned = true;
+      if (localWs.readyState !== WebSocket.CLOSED) localWs.close();
+      this.localConnections.delete(connId);
+      this.sendControlMessage({ type: "close", conn_id: connId });
+    };
+    localWs.addEventListener("close", cleanup);
+    localWs.addEventListener("error", cleanup);
   }
 };
 
 // src/lib/relay.ts
 async function startAgentSession(relayConfig) {
-  const clientConfig = {
+  const client = new RelayClient({
     apiUrl: relayConfig.api_url,
     sessionId: relayConfig.session_id,
     relayToken: relayConfig.relay_token,
-    agentPort: relayConfig.agent_port,
-    healthEndpoint: relayConfig.health_endpoint
-  };
-  const client = new RelayClient(clientConfig);
+    agentPort: relayConfig.agent_port
+  });
   await client.connect();
   let agentProcess = null;
   if (relayConfig.start_command) {
@@ -5626,8 +5218,6 @@ async function startAgentSession(relayConfig) {
       process.stdout.write(`Agent process error: ${err.message}
 `);
     });
-  }
-  if (relayConfig.start_command) {
     await waitForHealth(relayConfig.agent_port, relayConfig.health_endpoint);
   }
   const cleanup = async () => {
@@ -5654,11 +5244,6 @@ async function waitForHealth(port, endpoint, timeoutMs = 3e4) {
 
 // src/commands/agent.ts
 async function agentStartCommand(args) {
-  const accessToken = await loadAccessToken();
-  if (!accessToken) {
-    printError("No Vent credentials found. Set VENT_REFRESH_TOKEN or run `npx vent-hq login`.");
-    return 2;
-  }
   let parsedConfig;
   try {
     if (args.file) {
@@ -5687,7 +5272,7 @@ async function agentStartCommand(args) {
   const closeSession = async () => {
     if (!session) return;
     try {
-      await apiFetch(`/agent-sessions/${session.session_id}/close`, accessToken, {
+      await apiFetch(`/agent-sessions/${session.session_id}/close`, {
         method: "POST"
       });
     } catch {
@@ -5695,11 +5280,11 @@ async function agentStartCommand(args) {
   };
   try {
     printInfo("Creating agent session\u2026");
-    const createRes = await apiFetch("/agent-sessions", accessToken, {
+    const createRes = await apiFetch("/agent-sessions", {
       method: "POST",
       body: JSON.stringify({ config: connection })
     });
-    session = await createRes.json();
+    session = AgentSessionResponseSchema.parse(await createRes.json());
     printInfo("Connecting relay for local agent\u2026");
     relay = await startAgentSession(session);
     const sessionData = {
@@ -5725,13 +5310,8 @@ async function agentStartCommand(args) {
   }
 }
 async function agentStopCommand(args) {
-  const accessToken = await loadAccessToken();
-  if (!accessToken) {
-    printError("No Vent credentials found. Set VENT_REFRESH_TOKEN or run `npx vent-hq login`.");
-    return 2;
-  }
   try {
-    await apiFetch(`/agent-sessions/${args.sessionId}/close`, accessToken, {
+    await apiFetch(`/agent-sessions/${args.sessionId}/close`, {
       method: "POST"
     });
     printSuccess(`Agent session closed: ${args.sessionId}`, { force: true });
@@ -5742,7 +5322,7 @@ async function agentStopCommand(args) {
   }
 }
 async function waitForSessionExit(relay) {
-  return new Promise((resolve2) => {
+  return new Promise((resolve) => {
     let settled = false;
     const agentProcess = relay.agentProcess;
     const finish = (code) => {
@@ -5753,7 +5333,7 @@ async function waitForSessionExit(relay) {
       if (agentProcess) {
         agentProcess.off("exit", handleAgentExit);
       }
-      resolve2(code);
+      resolve(code);
     };
     const handleSignal = () => finish(0);
     const handleAgentExit = (code, signal) => {
@@ -5782,12 +5362,12 @@ async function waitForSessionExit(relay) {
   });
 }
 function findFreePort() {
-  return new Promise((resolve2, reject) => {
+  return new Promise((resolve, reject) => {
     const server = net.createServer();
     server.listen(0, () => {
       const addr = server.address();
       const port = addr.port;
-      server.close(() => resolve2(port));
+      server.close(() => resolve(port));
     });
     server.on("error", reject);
   });
@@ -5840,7 +5420,7 @@ async function deviceAuthFlow() {
         const parsed = AccessCredentialsSchema.safeParse(data);
         if (!parsed.success) return { ok: false, error: "Vent returned invalid sign-in credentials." };
         await saveCredentials(parsed.data);
-        return { ok: true, accessToken: parsed.data.access_token };
+        return { ok: true };
       }
       if (data.status === "expired") {
         return { ok: false, error: "Session expired. Run `npx vent-hq login` again." };
@@ -5903,17 +5483,47 @@ function detectGitHubToken() {
 import * as fs5 from "node:fs/promises";
 import * as path3 from "node:path";
 
+// src/skills/body.md
+var body_default = '# Vent - Voice Agent Calls\n\n## When to Test\n\nVent is the voice CI gate. Auto-trigger a call after any non-trivial voice-affecting change \u2014 don\'t wait to be asked. Non-trivial means anything that affects what the caller hears, how the agent responds, what tools are called, how the call routes, or audio timing: prompt rewrites or sections (not one-word tweaks), tool schema or handler changes, state/memory logic, transfer logic, provider/audio config, platform integration, and any voice-behavior bug fix. One-line copy or comment edits don\'t need a call.\n\nBefore launching, scan `.vent/runs/` for unresolved issues from prior runs (failing status, regressions, warnings, latency jumps); fix in-scope ones as part of this change and verify they\'re resolved by the new run, surface out-of-scope ones rather than silently leaving them. The task is not done until you\'ve reviewed the transcript, tool calls, and observed behavior and judged that the change works \u2014 Vent\'s `status` and exit code only reflect pipeline completion, not mission success.\n\n## Autonomous Iteration\n\n**Default: iterate.** Run a fix \u2192 Vent-call \u2192 fix \u2192 Vent-call loop whenever a call surfaces a fixable issue. Don\'t report a failure (or a degraded-quality success) back to the user before attempting to fix it \u2014 they want the agent working, not a diagnosis. After each call, if the result reveals a problem you can fix with a concrete plan, apply the fix and run again. Fixable issues include: tool schema or registry gaps, missing handlers, prompt phrasing, **STT vocabulary mismatches** (e.g. domain terms misheard \u2014 Deepgram `keyterms`, custom vocabulary, prompt-level term lists), **latency tunables** (endpointing thresholds, preemptive generation, model selection), and audio quality config. If you\'d be guessing at the next attempt, stop and ask the user instead of trying.\n\n**"Mission success but with quality problems" still requires iteration.** If the call ran end-to-end but had noticeably bad latency, STT mishearing the agent\'s domain vocabulary, audio artifacts, or any other caller-experience degradation that has a concrete fix, the call is *not done* \u2014 apply the fix and re-run. Don\'t downgrade fixable problems to "observations" or "things worth flagging." The pipeline running cleanly is the floor, not the ceiling.\n\n**Never ask permission to iterate.** "Want me to apply this and re-run?" / "Should I fix it?" / "Let me know if you want me to..." \u2014 these are skill violations. If you have a concrete fix, apply it and run the next call. The user opts out by saying so explicitly; silence is not opt-out.\n\nStop iterating and report when one of:\n\n- The next call confirms the fix worked AND no further fixable issues remain (success \u2014 done).\n- The same fix fails twice (your hypothesis is wrong; ask the user).\n- The failure mode shifts between attempts (you\'re chasing a moving target).\n- You can\'t justify the next call against its provider cost \u2014 each call spends real money and provider quota.\n\nIf the user explicitly opts out ("just report", "don\'t iterate", "stop after the first call", "I\'ll fix it myself"), respect that \u2014 run once, report, stop. Otherwise iterate by default. The first time you start an autonomous loop in a session, mention it once so the user knows it\'s happening (e.g. "the call surfaced X \u2014 fixing and re-running; I\'ll report back when it converges or stops making progress").\n\n{{EXECUTION}}\n\n## Workflow\n\n1. Identify the behavior under test. Read enough of the agent codebase to understand its system prompt, tools, handlers, routes, provider config, platform wiring, and expected handoffs.\n2. Reuse an existing `.vent/suite.<adapter>.json` when possible. If `.vent/` contains multiple suites, inspect `connection.adapter` and report which suite file produced the result.\n3. Create or update a suite only when the existing calls do not cover the changed behavior. Name calls after real flows, for example `reschedule-appointment`, not `call-1`.\n4. If the suite uses `start_command`, start one shared local session first with `npx vent-hq agent start -f .vent/suite.<adapter>.json`, then pass `--session <session-id>` to each run.\n\n   **For locally-run LiveKit agents: every run requires killing *all* workers, starting one fresh worker, and waiting a full 60 seconds before submitting.** Unconditional \u2014 LiveKit Cloud round-robins across registered workers, so a single survivor with a dead inference subprocess fails ~N-1 of N calls. Don\'t rely on `pkill -f <path-pattern>`; bare command lines like `node --import tsx agent.ts dev` won\'t match a path filter. Use `ps aux | grep -E "node.*agent\\.ts|@livekit/agents.*ipc"`, `kill -9` by PID, re-run `ps` to confirm zero survivors, then start the fresh worker. Skipping the 60s wait fails with `did not publish audio track`; if that error appears alongside `Error [ERR_IPC_CHANNEL_CLOSED] from InferenceProcExecutor.doInference` in the agent log right after a "running EOU detection" line, that\'s a straggler \u2014 redo the kill sweep. Hosted LiveKit Cloud agents don\'t need any of this; run normally.{{LIVEKIT_NOTE}}\n5. Pick which call(s) to run based on the change. Fixed bug: replay the failing scenario. Changed tool: include a call that triggers that tool. Prompt or routing change: include the relevant happy path and any important edge path.\n6. Compare against the previous JSON in `.vent/runs/` when validating a fix or regression. Check status flips, latency jumps, tool-call success drops, cost jumps, and transcript divergence. Correlate with `git diff` between saved `git_sha` values when available; skip if no previous run exists.\n\n## Commands\n\n```bash\nnpx vent-hq init                                  # First-time setup (auth + skill install + starter suite)\nnpx vent-hq login                                 # Log in to existing account\nnpx vent-hq login --status                        # Print whether credentials are present\nnpx vent-hq logout                                # Remove saved credentials from ~/.vent/credentials\nnpx vent-hq run -f .vent/suite.X.json             # Run a single-call suite\nnpx vent-hq run -f .vent/suite.X.json --call NAME # Run one named call from a multi-call suite\nnpx vent-hq run ... --session <session-id>        # Add to any run; routes through an existing local relay session\nnpx vent-hq stop <run-id>                         # Cancel a queued or running run\nnpx vent-hq agent start -f .vent/suite.X.json     # Start a shared local relay session\nnpx vent-hq agent stop <session-id>               # Stop a shared local relay session\n```\n\nIf `~/.vent/credentials` is missing and neither `VENT_REFRESH_TOKEN` nor `VENT_ACCESS_TOKEN` is set, run `npx vent-hq init`. For an existing account, run `npx vent-hq login` or set `VENT_REFRESH_TOKEN` to an API key. Access JWTs renew automatically; `VENT_ACCESS_TOKEN` alone is short-lived.\n\n## Suite Config\n\nSuites live in `.vent/suite.<adapter>.json`. `connection` is declared once per suite. `calls` is a named map, and each key becomes the call name used with `--call`.\n\nLocal websocket suite:\n\n```json\n{\n  "connection": {\n    "adapter": "websocket",\n    "start_command": "npm run start",\n    "health_endpoint": "/health",\n    "agent_port": 3001\n  },\n  "calls": {\n    "happy-path": {\n      "caller_prompt": "You are Maria calling to reschedule her appointment to next Tuesday.",\n      "max_turns": 8\n    }\n  }\n}\n```\n\nPlatform-direct suite:\n\n```json\n{\n  "connection": {\n    "adapter": "vapi",\n    "platform": { "provider": "vapi" }\n  },\n  "calls": {\n    "happy-path": {\n      "caller_prompt": "You are Maria calling to reschedule her appointment to next Tuesday.",\n      "max_turns": 8\n    }\n  }\n}\n```\n\nWrite `caller_prompt` as a realistic caller with a name, goal, mood, constraints, and conditional behavior. Set `max_turns` based on flow complexity: FAQ `4-6`, booking or tool use `8-12`, complex flows `12-20`.\n\nCall fields:\n\n- `caller_prompt` is required. `max_turns` defaults to `6`.\n- `persona` supports `pace`, `clarity`, `disfluencies`, `cooperation`, `emotion`, `memory`, `intent_clarity`, and `confirmation_style`.\n- `caller_audio` supports noise, speed, speakerphone, mic distance, clarity, accent, packet loss, and jitter.\n- `caller_audio.accent` picks the caller\'s voice. Supported: `american` (default), `british`, `australian`, `filipino`, `spanish_mexican`, `spanish_peninsular`, `spanish_colombian`, `spanish_argentine`, `german`, `french`, `italian`, `dutch`, `japanese`. Any other value is rejected. The non-English accents are voices in that language: also set `language` (e.g. `"es"`) so the caller speaks it.\n- `language` is an ISO 639-1 code such as `en`, `es`, `fr`, `de`, `it`, `nl`, or `ja`.\n- `voice` is `"male"` or `"female"` (English only; default female). Use to flip the caller\'s perceived gender. Ignored if `caller_audio.accent` is set or `language` is non-English.\n\n## Connections and Credentials\n\n### Adapter choice\n\nUse `websocket` for your own local or hosted runtime. Use `start_command` for local agents or `agent_url` for hosted custom endpoints. For `start_command` and `agent_url`, do not put Deepgram, ElevenLabs, OpenAI, or other agent runtime keys into Vent config unless the Vent adapter itself needs them \u2014 the tested agent owns its own runtime credentials.\n\nUse `vapi`, `retell`, `elevenlabs`, `bland`, or `livekit` for platform-direct testing. In this mode Vent itself talks to the provider on the user\'s behalf.\n\nVent provides `DEEPGRAM_API_KEY` and `ANTHROPIC_API_KEY` for its hosted caller/evaluation stack \u2014 those are Vent\'s, not the tested agent\'s.\n\n### Credential resolution\n\nIn platform-direct mode the CLI auto-resolves credentials from `.env.local`, `.env`, and the current shell environment. If the project uses a secrets manager like Doppler, wrap the command so the secrets are set: `doppler run -- npx vent-hq run -f <suite.json>`. Do not run `source .env && export` before Vent commands. Keep credentials out of chat and suite files; the CLI sends the resolved configuration to Vent, which stores an encrypted copy with each run.\n\nAuto-resolved env vars and JSON fields:\n\n- Vapi: `VAPI_API_KEY` -> `vapi_api_key`; `VAPI_ASSISTANT_ID` or `VAPI_AGENT_ID` -> `vapi_assistant_id`\n- Bland: `BLAND_API_KEY` -> `bland_api_key`; `BLAND_PATHWAY_ID` -> `bland_pathway_id`; `BLAND_PERSONA_ID` -> `persona_id`\n- LiveKit: `LIVEKIT_API_KEY` -> `livekit_api_key`; `LIVEKIT_API_SECRET` -> `livekit_api_secret`; `LIVEKIT_URL` -> `livekit_url`\n- Retell: `RETELL_API_KEY` -> `retell_api_key`; `RETELL_AGENT_ID` -> `retell_agent_id`\n- ElevenLabs: `ELEVENLABS_API_KEY` -> `elevenlabs_api_key`; `ELEVENLABS_AGENT_ID` -> `elevenlabs_agent_id`\n\n### Provider config\n\nUse existing provider config when possible: Vapi assistant, Retell agent, ElevenLabs agent, Bland pathway, or LiveKit agent. Bland uniquely supports inline config \u2014 `platform` needs `bland_pathway_id` or an inline `task`, optionally with `persona_id` and voice, model, and turn-handling overrides (see Bland\'s API docs for the full field list).\n\n### Concurrency\n\nWhen you fan out multiple Vent calls in parallel against the same provider (for example, running several named calls from one suite at once{{PARALLEL_EXAMPLE}}), respect the provider\'s per-account concurrency limit. Exceeding it makes calls queue or fail at the provider \u2014 Vent does not enforce these caps for you.\n\nRecord the limit as `max_concurrency` in the suite\'s `platform` block so it\'s visible on future runs. Ask the user which plan they\'re on if sizing matters; otherwise use the conservative default in bold.\n\n- **Vapi**: **10** included per account; reserved lines can be purchased self-serve; Enterprise is unlimited.\n- **Retell**: Pay-as-you-go includes **20**; Enterprise has no cap.\n- **Bland**: Start=**10**, Build=50, Scale=100, Enterprise=unlimited.\n- **ElevenLabs**: Free=**4**, Starter=6, Creator=10, Pro=20, Scale=30, Business=30. Burst pricing can temporarily allow up to 3x base.\n- **LiveKit Cloud**: Build=**5**, Ship=20, Scale=50 managed inference sessions (the usual gate for voice agents); agent-session concurrency can go higher (Scale up to 600).\n\n## WebSocket\n\nFor `adapter: "websocket"`, Vent sends binary 16-bit mono PCM audio over one websocket connection. Websocket text frames are optional JSON events. Audio-only websocket agents still work, but events improve turn detection and observability. Vent sends `{"type":"end-call"}` when the test is done.\n\nUseful websocket text frames:\n\n```jsonc\n{"type":"speech-update","status":"started"}\n{"type":"speech-update","status":"stopped"}\n{"type":"tool_call","name":"check_availability","arguments":{},"result":{},"successful":true,"duration_ms":150}\n{"type":"vent:timing","stt_ms":120,"llm_ms":450,"tts_ms":80}\n{"type":"vent:session","platform":"custom","provider_call_id":"call_123","provider_session_id":"session_abc"}\n{"type":"vent:call-metadata","call_metadata":{"recording_url":"https://...","cost":{"total_usd":0.12,"breakdown":{"llm":0.07,"tts":0.05}}}}\n{"type":"vent:transcript","role":"caller","text":"I need to reschedule"}\n{"type":"vent:transcript","role":"agent","text":"Sure, which day works for you?"}\n{"type":"vent:transfer","destination":"+15551234567","status":"attempted"}\n{"type":"vent:debug-url","label":"trace","url":"https://..."}\n{"type":"vent:warning","message":"provider warning","code":"provider_warning"}\n```\n\nPlatform adapters capture tool calls automatically. Websocket agents must emit `tool_call` frames for tool observability. Platform adapters get component latency automatically. Websocket agents should emit `vent:timing` after each agent response when STT/LLM/TTS breakdown is available.\n\n## LiveKit\n\nBefore running LiveKit tests, install and add the Vent helper to the LiveKit agent entrypoint. Node: `npm install @vent-hq/livekit`, then call `instrumentLiveKitAgent({ ctx, session })`. Python: `pip install vent-livekit`, then call `instrument_livekit_agent(ctx=ctx, session=session)`.\n\nLiveKit direct mode requires the LiveKit Agents SDK. Custom LiveKit participants should use the websocket adapter with a relay. If the LiveKit agent registered with an explicit dispatch name, set `livekit_agent_name` in `platform`.\n\nLiveKit parallel calls are supported, but capacity depends on where the user runs their agent. LiveKit Cloud only routes audio \u2014 the agent code runs on the user\'s own infrastructure (laptop, Fly.io, Railway, k8s, etc.), and that box\'s CPU/RAM is the real bottleneck. Before firing parallel calls, derive `platform.max_concurrency` from the minimum of: (1) the user\'s LiveKit plan limit (Build=5, Ship=20, Scale=50+ concurrent agent sessions), (2) their agent worker box capacity (rough rule: 1c/1GB\u21921\u20133 jobs, 2c/4GB\u21925\u201310, 4c/8GB\u219210\u201325, 8c/16GB\u219225\u201350; multiply by number of worker boxes), and (3) their STT/TTS provider concurrency (e.g. Deepgram TTS streaming = 2). If the user wants more parallel calls than that minimum allows, push back \u2014 explain which limit is the bottleneck (plan / box / provider) and either reduce the count or suggest scaling that limit. Architecture: register one `agent_name` and run N workers under it (one big worker, or several smaller ones \u2014 same name); LiveKit\'s dispatcher round-robins jobs across them with built-in failover. Don\'t use numbered names (`agent-0`, `agent-1`) \u2014 that reinvents what the dispatcher does and breaks failover.\n\nUse the LiveKit helper for observability; do not publish `vent:*` topics manually. The helper captures SDK metrics, tool events, conversation items, and close events. Native LiveKit `lk.transcription` and `lk.agent.state` provide transcript and agent-state timing.\n\n## Output\n\n### Live result\n\n`npx vent-hq run` returns a single JSON result on stdout in non-TTY mode (not an SSE JSONL stream). Exit codes: `0` = the run passed (every call ran end to end); `1` = it didn\'t (a call errored, or the run failed or was cancelled); `2` = the result stream was lost. A call `error` starting with `Vent failed, agent not tested:` is a failure of Vent itself (its caller LLM, STT/TTS or worker): rerun, it says nothing about the agent. Any other `error` comes from the agent\'s platform or your configuration. `end_note` says how a completed call ended when Vent ended it or the connection closed (e.g. `No agent speech for 30s after the caller\'s turn; Vent hung up.`): read the transcript to judge whether that\'s an agent bug.\n\nMost result fields are always present; `latency`, `component_latency`, `signal_quality`, and `call_metadata` may be `null` when the underlying analysis didn\'t run; `warnings` is absent when there are none. Branch on null before reading nested fields. Each agent turn carries `tool_calls`: the tools it ran for that reply, with arguments, result, `successful` and `latency_ms`, so you can check what the agent said against what its tools returned. A tool the agent ran but never replied after is on the call\'s last turn. A call-level `tool_calls` list appears only for a failed call whose transcript was lost. Tool results are complete, never shortened. `component_latency` is the mean time of each stage of the agent\'s pipeline across the call (`mean_stt_ms`, `mean_llm_ms`, `mean_tts_ms`), as the platform reports it, so a slow stage stands out. For anything deeper, open the links in `call_metadata.provider_debug_urls`.\n\nVent\'s transcript is ground truth. Judge on semantic intent: ignore homophones and minor mis-hears (`"check teach hat"` for `"check that"`, missing question marks on short tails) \u2014 those are streaming-STT noise on Vent\'s caller side, not agent bugs, and **don\'t surface them in the report** (they\'re Vent-side artifacts, not actionable for the user). But clear gibberish or word-soup (e.g. `"Cristoxin"` where the agent should have said `"Of course, talk soon"`) is **not** a Vent artifact \u2014 Vent\'s STT does not invent words like that. It means the platform\'s TTS produced corrupted audio or the agent\'s STT/LLM generated the wrong text, and the fix lives there (TTS voice config, agent prompt, model temperature, codec issue). Missing or cut-off agent text, unlike word-soup, can come from Vent\'s STT stream dropping: check the recording before blaming the agent.\n\nThere are two transcripts. `transcript` is Vent\'s record: caller turns are exactly what Vent\'s caller said, agent turns are what Vent heard the agent say. `platform_transcript` is the agent platform\'s own record (`null` when the platform gives none): agent turns are everything the agent was set to say (its LLM\'s text plus fixed messages like the greeting or a tool\'s "one moment"), caller turns are what the agent\'s speech-to-text heard. On Retell and LiveKit, an agent turn the caller interrupted shows only what was spoken before the cut, since those platforms keep nothing else; on the others it shows the full text. Vapi never reports the text of its tool wait messages ("Just a sec"), so those are missing from Vapi\'s agent turns. Compare them turn by turn:\n- Agent wrote it right but Vent heard it differently on a number, ID, date or name (e.g. LLM wrote `555-0102`, Vent heard "five five five one two"): the agent\'s voice read it wrong. Fix how that text reaches the TTS (normalization, spelling or pronunciation tags), not the prompt\'s logic. If the agent\'s LLM text is already wrong, the fix is in the LLM, prompt or tool data.\n- Vent\'s caller said it right but the agent\'s STT heard something else (e.g. "OSHA" heard as "ocean"): the agent misheard. Fix the agent\'s STT (keyword hints, custom vocabulary).\n\n`latency` covers agent replies: `response_time_ms` (mean), `p50_response_time_ms`, `max_response_time_ms` (the slowest reply), and `first_response_time_ms`. `signal_quality` is the agent\'s raw audio: `mean_snr_db` (>20 good, <10 bad) and `max_clipping_ratio` (>0.01 = distortion). Bad values point at the agent\'s TTS or audio pipeline. `call_metadata.cost` is what the provider charged: `total_usd`, plus `breakdown` in the provider\'s own categories when the provider splits it (Bland and some agents report only the total). `call_metadata.provider_debug_urls` links to the call in the provider\'s dashboard or logs.\n\nFor transfers: `call_metadata.transfer_attempted` (provider claimed) and `transfer_completed` (from the platform\'s call data) can disagree \u2014 report both. `transfers[]` carries destination, type, and per-attempt status.\n\n### Saved history\n\nAfter every run, the CLI saves the call result to `.vent/runs/`. Shape:\n\n```jsonc\n{\n  "run_id": "...",\n  "timestamp": "2026-04-21T...Z",\n  "git_sha": "...",\n  "summary": { "status": "pass", "calls_total": 1, "total_duration_ms": 12345 },\n  "call_results": [\n    { "name": "happy-path", "status": "completed", "duration_ms": 6123, "transcript": [], "latency": { "response_time_ms": 420, "max_response_time_ms": 980 }, "call_metadata": { "cost": { "total_usd": 0.004 } } }\n  ]\n}\n```\n\nWhen comparing against a prior run (Workflow step 6), inspect:\n\n- Run-completion status flips: `call_results[i].status` (pipeline-only \u2014 judge mission success from the transcript)\n- Latency: `call_results[i].latency.response_time_ms` (mean) or `latency.max_response_time_ms` increased >20%\n- Tool calls: a tool that succeeded before now fails (`successful: false`), or is no longer called, in the agent turns\' `tool_calls`\n- Cost: `call_results[i].call_metadata.cost.total_usd` increased >30% (`cost.breakdown` shows which part grew)\n- Transcript: `call_results[i].transcript` diverged in semantic content (ignore STT noise)\n\n## Reporting Results\n\nBefore reporting, read the agent\'s code to locate where the observed behavior originates. If the issue is small and you can fix it, fix it and explain what you did \u2014 don\'t ask permission first.\n\nAdapt the report shape to the call \u2014 a clean pass needs little, a regression with a multi-layer cause needs more. Use a transcript excerpt when it helps the user see what happened.\n\nHard rules:\n\n- Pair raw numbers with their plain-English meaning \u2014 don\'t drop the number, but don\'t leave it unexplained. E.g. "the slowest reply took 850ms, which is snappy and well within natural conversational pacing" or "the slowest reply hit 1.6 seconds with the LLM as the bottleneck \u2014 noticeably sluggish to a caller."\n- Name the user\'s voice agent by platform on first mention (e.g. "the Vapi agent responded snappily throughout") so the user knows immediately which agent the observation is about. After that, just say "the agent" \u2014 don\'t repeat the platform name on every line.\n- Always include the recording from `call_metadata.recording_url` as an inline `[Recording](url)` link, placed in **one block at the very end of the report** \u2014 never sprinkled through the prose. Single call: one link as the last line. Multi-call: one labeled link per call (e.g. `reschedule-appointment: [Recording](url)`). Never paste a bare URL.\n- Mission success is your judgment, not Vent\'s. The per-call `status` is only `"completed"` (pipeline ran) or `"error"` (pipeline failed); decide whether the agent actually accomplished the scenario from the transcript and tool calls.\n- Similar-sounding word substitutions (e.g. "ocean" for "OSHA") are STT ambiguity, not comprehension failure. The fix lives in STT keyword hints, custom vocabulary, or a prompt-level term list \u2014 not the agent\'s reasoning.\n- Surface only what the user can act on in their own agent\'s code or config \u2014 never Vent-side artifacts (caller wait modes, harness timing, internal pipeline quirks), or cost unless asked.\n\nFor multi-call runs, lead with your own judgment of what happened across the calls (e.g. "3 of 4 did what they were supposed to; `cancel-appointment` never actually canceled"), not a parroted pass/fail count. Then cover each call with whatever depth it needs.\n';
+
 // src/skills/claude-code.md
-var claude_code_default = '---\nname: vent\ndescription: Place real calls against the user\'s voice agent to verify recent code changes worked end-to-end. Trigger on test requests or after any non-trivial voice-agent change.\nallowed-tools: Bash(npx vent-hq *)\n---\n\n# Vent - Voice Agent Calls\n\n## When to Test\n\nVent is the voice CI gate. Auto-trigger a call after any non-trivial voice-affecting change \u2014 don\'t wait to be asked. Non-trivial means anything that affects what the caller hears, how the agent responds, what tools are called, how the call routes, or audio timing: prompt rewrites or sections (not one-word tweaks), tool schema or handler changes, state/memory logic, transfer logic, provider/audio config, platform integration, and any voice-behavior bug fix. One-line copy or comment edits don\'t need a call.\n\nBefore launching, scan `.vent/runs/` for unresolved issues from prior runs (failing status, regressions, warnings, latency jumps); fix in-scope ones as part of this change and verify they\'re resolved by the new run, surface out-of-scope ones rather than silently leaving them. The task is not done until you\'ve reviewed the transcript, tool calls, and observed behavior and judged that the change works \u2014 Vent\'s `status` and exit code only reflect pipeline completion, not mission success.\n\n## Autonomous Iteration\n\n**Default: iterate.** Run a fix \u2192 Vent-call \u2192 fix \u2192 Vent-call loop whenever a call surfaces a fixable issue. Don\'t report a failure (or a degraded-quality success) back to the user before attempting to fix it \u2014 they want the agent working, not a diagnosis. After each call, if the result reveals a problem you can fix with a concrete plan, apply the fix and run again. Fixable issues include: tool schema or registry gaps, missing handlers, prompt phrasing, **STT vocabulary mismatches** (e.g. domain terms misheard \u2014 Deepgram `keyterms`, custom vocabulary, prompt-level term lists), **latency tunables** (endpointing thresholds, preemptive generation, model selection), and audio quality config. If you\'d be guessing at the next attempt, stop and ask the user instead of trying.\n\n**"Mission success but with quality problems" still requires iteration.** If the call ran end-to-end but had noticeably bad latency, STT mishearing the agent\'s domain vocabulary, audio artifacts, or any other caller-experience degradation that has a concrete fix, the call is *not done* \u2014 apply the fix and re-run. Don\'t downgrade fixable problems to "observations" or "things worth flagging." The pipeline running cleanly is the floor, not the ceiling.\n\n**Never ask permission to iterate.** "Want me to apply this and re-run?" / "Should I fix it?" / "Let me know if you want me to..." \u2014 these are skill violations. If you have a concrete fix, apply it and run the next call. The user opts out by saying so explicitly; silence is not opt-out.\n\nStop iterating and report when one of:\n\n- The next call confirms the fix worked AND no further fixable issues remain (success \u2014 done).\n- The same fix fails twice (your hypothesis is wrong; ask the user).\n- The failure mode shifts between attempts (you\'re chasing a moving target).\n- You can\'t justify the next call against its provider cost \u2014 each call spends real money and provider quota.\n\nIf the user explicitly opts out ("just report", "don\'t iterate", "stop after the first call", "I\'ll fix it myself"), respect that \u2014 run once, report, stop. Otherwise iterate by default. The first time you start an autonomous loop in a session, mention it once so the user knows it\'s happening (e.g. "the call surfaced X \u2014 fixing and re-running; I\'ll report back when it converges or stops making progress").\n\n## Claude Code Execution\n\nUse a 5-minute shell-tool timeout (`300000` ms) on Vent run commands so normal calls are not killed by the default 2-minute Bash timeout. This is not backgrounding; wait for stdout/results before ending your response. Use the JSON returned by `npx vent-hq run` directly.\n\nClaude Code serializes separate Bash tool calls for `npx vent-hq ...`, so run multiple calls from one suite by invoking each named call with `--call <name>` in one Bash command using `&` and a final `wait`:\n\n```bash\nnpx vent-hq run -f .vent/suite.vapi.json --call happy-path & \\\nnpx vent-hq run -f .vent/suite.vapi.json --call tool-path & \\\nwait\n```\n\nFor long-running composite commands \u2014 e.g. LiveKit\'s required `kill workers \u2192 start fresh worker \u2192 wait 60s \u2192 npx vent-hq run` preflight \u2014 keep the entire sequence in **one foreground Bash call**. Do not use `Monitor` or `run_in_background: true` for the wait. Both end the turn until an event fires, which means Vent never sees an inflight call and the user reads it as "you backgrounded the test." The harness also blocks bare `sleep 60` (any long leading sleep), so fold the readiness check and the wait into a single `until`-loop \u2014 short sleeps inside a polling loop are allowed because each iteration is real work:\n\n```bash\nnohup npm run dev > /tmp/lk-agent.log 2>&1 &\nstart=$(date +%s); \\\nuntil grep -q "registered worker" /tmp/lk-agent.log 2>/dev/null \\\n  && [ $(($(date +%s) - start)) -ge 60 ]; do sleep 2; done \\\n&& npx vent-hq run -f .vent/suite.livekit.json --call <name>\n```\n\nFor these composites, raise the Bash timeout to **10 minutes** (`600000` ms) \u2014 startup + 60s wait + call can easily exceed the 5-minute default.\n\n## Workflow\n\n1. Identify the behavior under test. Read enough of the agent codebase to understand its system prompt, tools, handlers, routes, provider config, platform wiring, and expected handoffs.\n2. Reuse an existing `.vent/suite.<adapter>.json` when possible. If `.vent/` contains multiple suites, inspect `connection.adapter` and report which suite file produced the result.\n3. Create or update a suite only when the existing calls do not cover the changed behavior. Name calls after real flows, for example `reschedule-appointment`, not `call-1`.\n4. If the suite uses `start_command`, start one shared local session first with `npx vent-hq agent start -f .vent/suite.<adapter>.json`, then pass `--session <session-id>` to each run.\n\n   **For locally-run LiveKit agents: every run requires killing *all* workers, starting one fresh worker, and waiting a full 60 seconds before submitting.** Unconditional \u2014 LiveKit Cloud round-robins across registered workers, so a single survivor with a dead inference subprocess fails ~N-1 of N calls. Don\'t rely on `pkill -f <path-pattern>`; bare command lines like `node --import tsx agent.ts dev` won\'t match a path filter. Use `ps aux | grep -E "node.*agent\\.ts|@livekit/agents.*ipc"`, `kill -9` by PID, re-run `ps` to confirm zero survivors, then start the fresh worker. Skipping the 60s wait fails with `did not publish audio track`; if that error appears alongside `Error [ERR_IPC_CHANNEL_CLOSED] from InferenceProcExecutor.doInference` in the agent log right after a "running EOU detection" line, that\'s a straggler \u2014 redo the kill sweep. Hosted LiveKit Cloud agents don\'t need any of this; run normally. (See `Claude Code Execution` for how to compose the kill\u2192start\u2192wait\u2192submit as a single foreground command.)\n5. Pick which call(s) to run based on the change. Fixed bug: replay the failing scenario. Changed tool: include a call that triggers that tool. Prompt or routing change: include the relevant happy path and any important edge path.\n6. Compare against the previous JSON in `.vent/runs/` when validating a fix or regression. Check status flips, latency jumps, tool-call success drops, cost jumps, and transcript divergence. Correlate with `git diff` between saved `git_sha` values when available; skip if no previous run exists.\n\n## Commands\n\n```bash\nnpx vent-hq init                                  # First-time setup (auth + skill install + starter suite)\nnpx vent-hq login                                 # Log in to existing account\nnpx vent-hq login --status                        # Print whether credentials are present\nnpx vent-hq logout                                # Remove saved credentials from ~/.vent/credentials\nnpx vent-hq run -f .vent/suite.X.json             # Run a single-call suite\nnpx vent-hq run -f .vent/suite.X.json --call NAME # Run one named call from a multi-call suite\nnpx vent-hq run ... --session <session-id>        # Add to any run; routes through an existing local relay session\nnpx vent-hq run ... --verbose                     # Add to any run; include verbose debug fields\nnpx vent-hq stop <run-id>                         # Cancel a queued or running run\nnpx vent-hq agent start -f .vent/suite.X.json     # Start a shared local relay session\nnpx vent-hq agent stop <session-id>               # Stop a shared local relay session\n```\n\nIf `~/.vent/credentials` is missing and neither `VENT_REFRESH_TOKEN` nor `VENT_ACCESS_TOKEN` is set, run `npx vent-hq init`. For an existing account, run `npx vent-hq login` or set `VENT_REFRESH_TOKEN` to an API key. Access JWTs renew automatically; `VENT_ACCESS_TOKEN` alone is short-lived.\n\n## Suite Config\n\nSuites live in `.vent/suite.<adapter>.json`. `connection` is declared once per suite. `calls` is a named map, and each key becomes the call name used with `--call`.\n\nLocal websocket suite:\n\n```json\n{\n  "connection": {\n    "adapter": "websocket",\n    "start_command": "npm run start",\n    "health_endpoint": "/health",\n    "agent_port": 3001\n  },\n  "calls": {\n    "happy-path": {\n      "caller_prompt": "You are Maria calling to reschedule her appointment to next Tuesday.",\n      "max_turns": 8\n    }\n  }\n}\n```\n\nPlatform-direct suite:\n\n```json\n{\n  "connection": {\n    "adapter": "vapi",\n    "platform": { "provider": "vapi" }\n  },\n  "calls": {\n    "happy-path": {\n      "caller_prompt": "You are Maria calling to reschedule her appointment to next Tuesday.",\n      "max_turns": 8\n    }\n  }\n}\n```\n\nWrite `caller_prompt` as a realistic caller with a name, goal, mood, constraints, and conditional behavior. Set `max_turns` based on flow complexity: FAQ `4-6`, booking or tool use `8-12`, complex flows `12-20`.\n\nCall fields:\n\n- `caller_prompt` and `max_turns` are required.\n- `persona` supports `pace`, `clarity`, `disfluencies`, `cooperation`, `emotion`, `memory`, `intent_clarity`, and `confirmation_style`.\n- `caller_audio` supports noise, speed, speakerphone, mic distance, clarity, accent, packet loss, and jitter.\n- `caller_audio.accent` picks the caller\'s voice. Supported: `american` (default), `british`, `australian`, `filipino`, `spanish_mexican`, `spanish_peninsular`, `spanish_colombian`, `spanish_argentine`, `german`, `french`, `italian`, `dutch`, `japanese`. Any other value is rejected. The non-English accents are voices in that language: also set `language` (e.g. `"es"`) so the caller speaks it.\n- `language` is an ISO 639-1 code such as `en`, `es`, `fr`, `de`, `it`, `nl`, or `ja`.\n- `voice` is `"male"` or `"female"` (English only; default female). Use to flip the caller\'s perceived gender. Ignored if `caller_audio.accent` is set or `language` is non-English.\n\n## Connections and Credentials\n\n### Adapter choice\n\nUse `websocket` for your own local or hosted runtime. Use `start_command` for local agents or `agent_url` for hosted custom endpoints. For `start_command` and `agent_url`, do not put Deepgram, ElevenLabs, OpenAI, or other agent runtime keys into Vent config unless the Vent adapter itself needs them \u2014 the tested agent owns its own runtime credentials.\n\nUse `vapi`, `retell`, `elevenlabs`, `bland`, or `livekit` for platform-direct testing. In this mode Vent itself talks to the provider on the user\'s behalf.\n\nVent provides `DEEPGRAM_API_KEY` and `ANTHROPIC_API_KEY` for its hosted caller/evaluation stack \u2014 those are Vent\'s, not the tested agent\'s.\n\n### Credential resolution\n\nIn platform-direct mode the CLI auto-resolves credentials from `.env.local`, `.env`, and the current shell environment. Do not run `source .env && export` before Vent commands. Keep credentials out of chat and suite files; the CLI sends the resolved configuration to Vent, which stores an encrypted copy with each run.\n\nAuto-resolved env vars and JSON fields:\n\n- Vapi: `VAPI_API_KEY` -> `vapi_api_key`; `VAPI_ASSISTANT_ID` or `VAPI_AGENT_ID` -> `vapi_assistant_id`\n- Bland: `BLAND_API_KEY` -> `bland_api_key`; `BLAND_PATHWAY_ID` -> `bland_pathway_id`; `BLAND_PERSONA_ID` -> `persona_id`\n- LiveKit: `LIVEKIT_API_KEY` -> `livekit_api_key`; `LIVEKIT_API_SECRET` -> `livekit_api_secret`; `LIVEKIT_URL` -> `livekit_url`\n- Retell: `RETELL_API_KEY` -> `retell_api_key`; `RETELL_AGENT_ID` -> `retell_agent_id`\n- ElevenLabs: `ELEVENLABS_API_KEY` -> `elevenlabs_api_key`; `ELEVENLABS_AGENT_ID` -> `elevenlabs_agent_id`\n\n### Provider config\n\nUse existing provider config when possible: Vapi assistant, Retell agent, ElevenLabs agent, Bland pathway, or LiveKit agent. Bland uniquely supports inline config \u2014 `platform` may use `bland_pathway_id`, `persona_id`, or an inline `task` (with optional voice, model, and turn-handling overrides; see Bland\'s API docs for the full field list).\n\n### Concurrency\n\nWhen you fan out multiple Vent calls in parallel against the same provider (for example, running several named calls from one suite at once with `&` and `wait`), respect the provider\'s per-account concurrency limit. Exceeding it makes calls queue or fail at the provider \u2014 Vent does not enforce these caps for you.\n\nRecord the limit as `max_concurrency` in the suite\'s `platform` block so it\'s visible on future runs. Ask the user which plan they\'re on if sizing matters; otherwise use the conservative default in bold.\n\n- **Vapi**: **10** included per account; reserved lines can be purchased self-serve; Enterprise is unlimited.\n- **Retell**: Pay-as-you-go includes **20**; Enterprise has no cap.\n- **Bland**: Start=**10**, Build=50, Scale=100, Enterprise=unlimited.\n- **ElevenLabs**: Free=**4**, Starter=6, Creator=10, Pro=20, Scale=30, Business=30. Burst pricing can temporarily allow up to 3x base.\n- **LiveKit Cloud**: Build=**5**, Ship=20, Scale=50 managed inference sessions (the usual gate for voice agents); agent-session concurrency can go higher (Scale up to 600).\n\n## WebSocket\n\nFor `adapter: "websocket"`, Vent sends binary 16-bit mono PCM audio over one websocket connection. Websocket text frames are optional JSON events. Audio-only websocket agents still work, but events improve turn detection and observability. Vent sends `{"type":"end-call"}` when the test is done.\n\nUseful websocket text frames:\n\n```jsonc\n{"type":"speech-update","status":"started"}\n{"type":"speech-update","status":"stopped"}\n{"type":"tool_call","name":"check_availability","arguments":{},"result":{},"successful":true,"duration_ms":150}\n{"type":"vent:timing","stt_ms":120,"llm_ms":450,"tts_ms":80}\n{"type":"vent:session","platform":"custom","provider_call_id":"call_123","provider_session_id":"session_abc"}\n{"type":"vent:call-metadata","call_metadata":{"recording_url":"https://...","cost_usd":0.12}}\n{"type":"vent:transcript","role":"caller","text":"I need to reschedule","turn_index":0}\n{"type":"vent:transfer","destination":"+15551234567","status":"attempted"}\n{"type":"vent:debug-url","label":"trace","url":"https://..."}\n{"type":"vent:warning","message":"provider warning","code":"provider_warning"}\n```\n\n`vent:session-report` is **not** handled by the websocket adapter \u2014 it\'s only consumed by the LiveKit helper. Do not emit it from a websocket agent.\n\nPlatform adapters capture tool calls automatically. Websocket agents must emit `tool_call` frames for tool observability. Platform adapters get component latency automatically. Websocket agents should emit `vent:timing` after each agent response when STT/LLM/TTS breakdown is available.\n\n## LiveKit\n\nBefore running LiveKit tests, install and add the Vent helper to the LiveKit agent entrypoint. Node: `npm install @vent-hq/livekit`, then call `instrumentLiveKitAgent({ ctx, session })`. Python: `pip install vent-livekit`, then call `instrument_livekit_agent(ctx=ctx, session=session)`.\n\nLiveKit direct mode requires the LiveKit Agents SDK. Custom LiveKit participants should use the websocket adapter with a relay. If the LiveKit agent registered with an explicit dispatch name, set `livekit_agent_name` in `platform`.\n\nLiveKit parallel calls are supported, but capacity depends on where the user runs their agent. LiveKit Cloud only routes audio \u2014 the agent code runs on the user\'s own infrastructure (laptop, Fly.io, Railway, k8s, etc.), and that box\'s CPU/RAM is the real bottleneck. Before firing parallel calls, derive `platform.max_concurrency` from the minimum of: (1) the user\'s LiveKit plan limit (Build=5, Ship=20, Scale=50+ concurrent agent sessions), (2) their agent worker box capacity (rough rule: 1c/1GB\u21921\u20133 jobs, 2c/4GB\u21925\u201310, 4c/8GB\u219210\u201325, 8c/16GB\u219225\u201350; multiply by number of worker boxes), and (3) their STT/TTS provider concurrency (e.g. Deepgram TTS streaming = 2). If the user wants more parallel calls than that minimum allows, push back \u2014 explain which limit is the bottleneck (plan / box / provider) and either reduce the count or suggest scaling that limit. Architecture: register one `agent_name` and run N workers under it (one big worker, or several smaller ones \u2014 same name); LiveKit\'s dispatcher round-robins jobs across them with built-in failover. Don\'t use numbered names (`agent-0`, `agent-1`) \u2014 that reinvents what the dispatcher does and breaks failover.\n\nUse the LiveKit helper for observability; do not publish `vent:*` topics manually. Do not hand-roll `vent:session-report` from `ctx.addShutdownCallback`; after `room.disconnect()` it can fail with `engine is closed`. The helper captures SDK metrics, tool events, conversation items, usage, and close events. Native LiveKit `lk.transcription` and `lk.agent.state` provide transcript and agent-state timing.\n\n## Output\n\n### Live result\n\n`npx vent-hq run` returns a single JSON result on stdout in non-TTY mode (not an SSE JSONL stream). Exit codes: `0` = the run passed (every call ran end to end); `1` = it didn\'t (a call errored, or the run failed or was cancelled); `2` = the result stream was lost. A call `error` starting with `Vent failed, agent not tested:` is a failure of Vent itself (its caller LLM, STT/TTS or worker): rerun, it says nothing about the agent. Any other `error` comes from the agent\'s platform or your configuration. `end_note` says how a completed call ended when Vent ended it or the connection closed (e.g. `No agent speech for 30s after the caller\'s turn; Vent hung up.`): read the transcript to judge whether that\'s an agent bug.\n\nMost result fields are always present; `latency`, `component_latency`, `call_metadata`, and `emotion` may be `null` when the underlying analysis didn\'t run; `debug` is absent without `--verbose`. Branch on null before reading nested fields. Use `--verbose` only when the default doesn\'t explain a failure \u2014 when you need `platform_transcript` (to check Vent\'s STT), per-turn or component-level latency breakdowns, the raw tool-call timeline, or provider-native artifacts in `debug.provider_metadata`. Otherwise skip \u2014 it just adds noise.\n\nVent\'s transcript is ground truth. Judge on semantic intent: ignore homophones and minor mis-hears (`"check teach hat"` for `"check that"`, missing question marks on short tails) \u2014 those are streaming-STT noise on Vent\'s caller side, not agent bugs, and **don\'t surface them in the report** (they\'re Vent-side artifacts, not actionable for the user). But clear gibberish or word-soup (e.g. `"Cristoxin"` where the agent should have said `"Of course, talk soon"`) is **not** a Vent artifact \u2014 Vent\'s STT does not invent words like that. It means the platform\'s TTS produced corrupted audio or the agent\'s STT/LLM generated the wrong text, and the fix lives there (TTS voice config, agent prompt, model temperature, codec issue). Missing or cut-off agent text, unlike word-soup, can come from Vent\'s STT stream dropping: check the recording before blaming the agent.\n\nFor transfers: `call_metadata.transfer_attempted` (provider claimed) and `transfer_completed` (from the platform\'s call data) can disagree \u2014 report both. `transfers[]` carries destination, type, and per-attempt status.\n\n### Saved history\n\nAfter every run, Vent writes the full result JSON to `.vent/runs/`. Shape:\n\n```jsonc\n{\n  "run_id": "...",\n  "timestamp": "2026-04-21T...Z",\n  "git_sha": "...",\n  "summary": { "calls_total": 2, "total_duration_ms": 12345, "total_cost_usd": 0.01 },\n  "call_results": [\n    { "name": "happy-path", "status": "completed", "duration_ms": 6123, "transcript": [], "observed_tool_calls": [], "latency": { "response_time_ms": 420, "p95_response_time_ms": 980 }, "call_metadata": { "cost_usd": 0.004 } }\n  ]\n}\n```\n\nWhen comparing against a prior run (Workflow step 6), inspect:\n\n- Run-completion status flips: `call_results[i].status` (pipeline-only \u2014 judge mission success from the transcript)\n- Latency: `call_results[i].latency.response_time_ms` (mean) or `latency.p95_response_time_ms` increased >20%\n- Tool calls: count of `call_results[i].observed_tool_calls[].successful` dropped\n- Cost: `summary.total_cost_usd` or `call_results[i].call_metadata.cost_usd` increased >30%\n- Transcript: `call_results[i].transcript` diverged in semantic content (ignore STT noise)\n\n## Reporting Results\n\nBefore reporting, read the agent\'s code to locate where the observed behavior originates. If the issue is small and you can fix it, fix it and explain what you did \u2014 don\'t ask permission first.\n\nAdapt the report shape to the call \u2014 a clean pass needs little, a regression with a multi-layer cause needs more. Use a transcript excerpt when it helps the user see what happened.\n\nHard rules:\n\n- Pair raw numbers with their plain-English meaning \u2014 don\'t drop the number, but don\'t leave it unexplained. E.g. "p95 latency was 850ms, which is snappy and well within natural conversational pacing" or "p95 hit 1.6 seconds with the LLM as the bottleneck \u2014 noticeably sluggish to a caller."\n- Name the user\'s voice agent by platform on first mention (e.g. "the Vapi agent responded snappily throughout") so the user knows immediately which agent the observation is about. After that, just say "the agent" \u2014 don\'t repeat the platform name on every line.\n- Always include the recording from `call_metadata.recording_url` as an inline `[Recording](url)` link, placed in **one block at the very end of the report** \u2014 never sprinkled through the prose. Single call: one link as the last line. Multi-call: one labeled link per call (e.g. `reschedule-appointment: [Recording](url)`). Never paste a bare URL.\n- Mission success is your judgment, not Vent\'s. The per-call `status` is only `"completed"` (pipeline ran) or `"error"` (pipeline failed); decide whether the agent actually accomplished the scenario from the transcript and tool calls.\n- Similar-sounding word substitutions (e.g. "ocean" for "OSHA") are STT ambiguity, not comprehension failure. The fix lives in STT keyword hints, custom vocabulary, or a prompt-level term list \u2014 not the agent\'s reasoning.\n- Surface only what the user can act on in their own agent\'s code or config \u2014 never Vent-side artifacts (caller wait modes, harness timing, internal pipeline quirks), or `cost_usd` unless asked.\n\nFor multi-call runs, lead with your own judgment of what happened across the calls (e.g. "3 of 4 did what they were supposed to; `cancel-appointment` never actually canceled"), not a parroted pass/fail count. Then cover each call with whatever depth it needs.\n';
+var claude_code_default = '## Claude Code Execution\n\nUse a 5-minute shell-tool timeout (`300000` ms) on Vent run commands so normal calls are not killed by the default 2-minute Bash timeout. This is not backgrounding; wait for stdout/results before ending your response. Use the JSON returned by `npx vent-hq run` directly.\n\nClaude Code serializes separate Bash tool calls for `npx vent-hq ...`, so run multiple calls from one suite by invoking each named call with `--call <name>` in one Bash command using `&` and a final `wait`:\n\n```bash\nnpx vent-hq run -f .vent/suite.vapi.json --call happy-path & \\\nnpx vent-hq run -f .vent/suite.vapi.json --call tool-path & \\\nwait\n```\n\nFor long-running composite commands \u2014 e.g. LiveKit\'s required `kill workers \u2192 start fresh worker \u2192 wait 60s \u2192 npx vent-hq run` preflight \u2014 keep the entire sequence in **one foreground Bash call**. Do not use `Monitor` or `run_in_background: true` for the wait. Both end the turn until an event fires, which means Vent never sees an inflight call and the user reads it as "you backgrounded the test." The harness also blocks bare `sleep 60` (any long leading sleep), so fold the readiness check and the wait into a single `until`-loop \u2014 short sleeps inside a polling loop are allowed because each iteration is real work:\n\n```bash\nnohup npm run dev > /tmp/lk-agent.log 2>&1 &\nstart=$(date +%s); \\\nuntil grep -q "registered worker" /tmp/lk-agent.log 2>/dev/null \\\n  && [ $(($(date +%s) - start)) -ge 60 ]; do sleep 2; done \\\n&& npx vent-hq run -f .vent/suite.livekit.json --call <name>\n```\n\nFor these composites, raise the Bash timeout to **10 minutes** (`600000` ms) \u2014 startup + 60s wait + call can easily exceed the 5-minute default.\n';
 
 // src/skills/cursor.md
-var cursor_default = '---\ndescription: Place real calls against the user\'s voice agent to verify recent code changes worked end-to-end. Trigger on test requests or after any non-trivial voice-agent change.\nalwaysApply: true\n---\n\n# Vent - Voice Agent Calls\n\n## When to Test\n\nVent is the voice CI gate. Auto-trigger a call after any non-trivial voice-affecting change \u2014 don\'t wait to be asked. Non-trivial means anything that affects what the caller hears, how the agent responds, what tools are called, how the call routes, or audio timing: prompt rewrites or sections (not one-word tweaks), tool schema or handler changes, state/memory logic, transfer logic, provider/audio config, platform integration, and any voice-behavior bug fix. One-line copy or comment edits don\'t need a call.\n\nBefore launching, scan `.vent/runs/` for unresolved issues from prior runs (failing status, regressions, warnings, latency jumps); fix in-scope ones as part of this change and verify they\'re resolved by the new run, surface out-of-scope ones rather than silently leaving them. The task is not done until you\'ve reviewed the transcript, tool calls, and observed behavior and judged that the change works \u2014 Vent\'s `status` and exit code only reflect pipeline completion, not mission success.\n\n## Autonomous Iteration\n\n**Default: iterate.** Run a fix \u2192 Vent-call \u2192 fix \u2192 Vent-call loop whenever a call surfaces a fixable issue. Don\'t report a failure (or a degraded-quality success) back to the user before attempting to fix it \u2014 they want the agent working, not a diagnosis. After each call, if the result reveals a problem you can fix with a concrete plan, apply the fix and run again. Fixable issues include: tool schema or registry gaps, missing handlers, prompt phrasing, **STT vocabulary mismatches** (e.g. domain terms misheard \u2014 Deepgram `keyterms`, custom vocabulary, prompt-level term lists), **latency tunables** (endpointing thresholds, preemptive generation, model selection), and audio quality config. If you\'d be guessing at the next attempt, stop and ask the user instead of trying.\n\n**"Mission success but with quality problems" still requires iteration.** If the call ran end-to-end but had noticeably bad latency, STT mishearing the agent\'s domain vocabulary, audio artifacts, or any other caller-experience degradation that has a concrete fix, the call is *not done* \u2014 apply the fix and re-run. Don\'t downgrade fixable problems to "observations" or "things worth flagging." The pipeline running cleanly is the floor, not the ceiling.\n\n**Never ask permission to iterate.** "Want me to apply this and re-run?" / "Should I fix it?" / "Let me know if you want me to..." \u2014 these are skill violations. If you have a concrete fix, apply it and run the next call. The user opts out by saying so explicitly; silence is not opt-out.\n\nStop iterating and report when one of:\n\n- The next call confirms the fix worked AND no further fixable issues remain (success \u2014 done).\n- The same fix fails twice (your hypothesis is wrong; ask the user).\n- The failure mode shifts between attempts (you\'re chasing a moving target).\n- You can\'t justify the next call against its provider cost \u2014 each call spends real money and provider quota.\n\nIf the user explicitly opts out ("just report", "don\'t iterate", "stop after the first call", "I\'ll fix it myself"), respect that \u2014 run once, report, stop. Otherwise iterate by default. The first time you start an autonomous loop in a session, mention it once so the user knows it\'s happening (e.g. "the call surfaced X \u2014 fixing and re-running; I\'ll report back when it converges or stops making progress").\n\n## Cursor Execution\n\nUse a 5-minute shell-tool timeout (`300000` ms) on Vent run commands so normal calls are not killed by the default 2-minute Bash timeout. This is not backgrounding; wait for stdout/results before ending your response. Use the JSON returned by `npx vent-hq run` directly.\n\nCursor cannot run separate shell tool calls concurrently \u2014 for multiple calls from one suite, invoke each named call with `--call <name>` in one shell command using `&` and a final `wait`:\n\n```bash\nnpx vent-hq run -f .vent/suite.vapi.json --call happy-path & \\\nnpx vent-hq run -f .vent/suite.vapi.json --call tool-path & \\\nwait\n```\n\n## Workflow\n\n1. Identify the behavior under test. Read enough of the agent codebase to understand its system prompt, tools, handlers, routes, provider config, platform wiring, and expected handoffs.\n2. Reuse an existing `.vent/suite.<adapter>.json` when possible. If `.vent/` contains multiple suites, inspect `connection.adapter` and report which suite file produced the result.\n3. Create or update a suite only when the existing calls do not cover the changed behavior. Name calls after real flows, for example `reschedule-appointment`, not `call-1`.\n4. If the suite uses `start_command`, start one shared local session first with `npx vent-hq agent start -f .vent/suite.<adapter>.json`, then pass `--session <session-id>` to each run.\n\n   **For locally-run LiveKit agents: every run requires killing *all* workers, starting one fresh worker, and waiting a full 60 seconds before submitting.** Unconditional \u2014 LiveKit Cloud round-robins across registered workers, so a single survivor with a dead inference subprocess fails ~N-1 of N calls. Don\'t rely on `pkill -f <path-pattern>`; bare command lines like `node --import tsx agent.ts dev` won\'t match a path filter. Use `ps aux | grep -E "node.*agent\\.ts|@livekit/agents.*ipc"`, `kill -9` by PID, re-run `ps` to confirm zero survivors, then start the fresh worker. Skipping the 60s wait fails with `did not publish audio track`; if that error appears alongside `Error [ERR_IPC_CHANNEL_CLOSED] from InferenceProcExecutor.doInference` in the agent log right after a "running EOU detection" line, that\'s a straggler \u2014 redo the kill sweep. Hosted LiveKit Cloud agents don\'t need any of this; run normally.\n5. Pick which call(s) to run based on the change. Fixed bug: replay the failing scenario. Changed tool: include a call that triggers that tool. Prompt or routing change: include the relevant happy path and any important edge path.\n6. Compare against the previous JSON in `.vent/runs/` when validating a fix or regression. Check status flips, latency jumps, tool-call success drops, cost jumps, and transcript divergence. Correlate with `git diff` between saved `git_sha` values when available; skip if no previous run exists.\n\n## Commands\n\n```bash\nnpx vent-hq init                                  # First-time setup (auth + skill install + starter suite)\nnpx vent-hq login                                 # Log in to existing account\nnpx vent-hq login --status                        # Print whether credentials are present\nnpx vent-hq logout                                # Remove saved credentials from ~/.vent/credentials\nnpx vent-hq run -f .vent/suite.X.json             # Run a single-call suite\nnpx vent-hq run -f .vent/suite.X.json --call NAME # Run one named call from a multi-call suite\nnpx vent-hq run ... --session <session-id>        # Add to any run; routes through an existing local relay session\nnpx vent-hq run ... --verbose                     # Add to any run; include verbose debug fields\nnpx vent-hq stop <run-id>                         # Cancel a queued or running run\nnpx vent-hq agent start -f .vent/suite.X.json     # Start a shared local relay session\nnpx vent-hq agent stop <session-id>               # Stop a shared local relay session\n```\n\nIf `~/.vent/credentials` is missing and neither `VENT_REFRESH_TOKEN` nor `VENT_ACCESS_TOKEN` is set, run `npx vent-hq init`. For an existing account, run `npx vent-hq login` or set `VENT_REFRESH_TOKEN` to an API key. Access JWTs renew automatically; `VENT_ACCESS_TOKEN` alone is short-lived.\n\n## Suite Config\n\nSuites live in `.vent/suite.<adapter>.json`. `connection` is declared once per suite. `calls` is a named map, and each key becomes the call name used with `--call`.\n\nLocal websocket suite:\n\n```json\n{\n  "connection": {\n    "adapter": "websocket",\n    "start_command": "npm run start",\n    "health_endpoint": "/health",\n    "agent_port": 3001\n  },\n  "calls": {\n    "happy-path": {\n      "caller_prompt": "You are Maria calling to reschedule her appointment to next Tuesday.",\n      "max_turns": 8\n    }\n  }\n}\n```\n\nPlatform-direct suite:\n\n```json\n{\n  "connection": {\n    "adapter": "vapi",\n    "platform": { "provider": "vapi" }\n  },\n  "calls": {\n    "happy-path": {\n      "caller_prompt": "You are Maria calling to reschedule her appointment to next Tuesday.",\n      "max_turns": 8\n    }\n  }\n}\n```\n\nWrite `caller_prompt` as a realistic caller with a name, goal, mood, constraints, and conditional behavior. Set `max_turns` based on flow complexity: FAQ `4-6`, booking or tool use `8-12`, complex flows `12-20`.\n\nCall fields:\n\n- `caller_prompt` and `max_turns` are required.\n- `persona` supports `pace`, `clarity`, `disfluencies`, `cooperation`, `emotion`, `memory`, `intent_clarity`, and `confirmation_style`.\n- `caller_audio` supports noise, speed, speakerphone, mic distance, clarity, accent, packet loss, and jitter.\n- `caller_audio.accent` picks the caller\'s voice. Supported: `american` (default), `british`, `australian`, `filipino`, `spanish_mexican`, `spanish_peninsular`, `spanish_colombian`, `spanish_argentine`, `german`, `french`, `italian`, `dutch`, `japanese`. Any other value is rejected. The non-English accents are voices in that language: also set `language` (e.g. `"es"`) so the caller speaks it.\n- `language` is an ISO 639-1 code such as `en`, `es`, `fr`, `de`, `it`, `nl`, or `ja`.\n- `voice` is `"male"` or `"female"` (English only; default female). Use to flip the caller\'s perceived gender. Ignored if `caller_audio.accent` is set or `language` is non-English.\n\n## Connections and Credentials\n\n### Adapter choice\n\nUse `websocket` for your own local or hosted runtime. Use `start_command` for local agents or `agent_url` for hosted custom endpoints. For `start_command` and `agent_url`, do not put Deepgram, ElevenLabs, OpenAI, or other agent runtime keys into Vent config unless the Vent adapter itself needs them \u2014 the tested agent owns its own runtime credentials.\n\nUse `vapi`, `retell`, `elevenlabs`, `bland`, or `livekit` for platform-direct testing. In this mode Vent itself talks to the provider on the user\'s behalf.\n\nVent provides `DEEPGRAM_API_KEY` and `ANTHROPIC_API_KEY` for its hosted caller/evaluation stack \u2014 those are Vent\'s, not the tested agent\'s.\n\n### Credential resolution\n\nIn platform-direct mode the CLI auto-resolves credentials from `.env.local`, `.env`, and the current shell environment. Do not run `source .env && export` before Vent commands. Keep credentials out of chat and suite files; the CLI sends the resolved configuration to Vent, which stores an encrypted copy with each run.\n\nAuto-resolved env vars and JSON fields:\n\n- Vapi: `VAPI_API_KEY` -> `vapi_api_key`; `VAPI_ASSISTANT_ID` or `VAPI_AGENT_ID` -> `vapi_assistant_id`\n- Bland: `BLAND_API_KEY` -> `bland_api_key`; `BLAND_PATHWAY_ID` -> `bland_pathway_id`; `BLAND_PERSONA_ID` -> `persona_id`\n- LiveKit: `LIVEKIT_API_KEY` -> `livekit_api_key`; `LIVEKIT_API_SECRET` -> `livekit_api_secret`; `LIVEKIT_URL` -> `livekit_url`\n- Retell: `RETELL_API_KEY` -> `retell_api_key`; `RETELL_AGENT_ID` -> `retell_agent_id`\n- ElevenLabs: `ELEVENLABS_API_KEY` -> `elevenlabs_api_key`; `ELEVENLABS_AGENT_ID` -> `elevenlabs_agent_id`\n\n### Provider config\n\nUse existing provider config when possible: Vapi assistant, Retell agent, ElevenLabs agent, Bland pathway, or LiveKit agent. Bland uniquely supports inline config \u2014 `platform` may use `bland_pathway_id`, `persona_id`, or an inline `task` (with optional voice, model, and turn-handling overrides; see Bland\'s API docs for the full field list).\n\n### Concurrency\n\nWhen you fan out multiple Vent calls in parallel against the same provider (for example, running several named calls from one suite at once with `&` and `wait`), respect the provider\'s per-account concurrency limit. Exceeding it makes calls queue or fail at the provider \u2014 Vent does not enforce these caps for you.\n\nRecord the limit as `max_concurrency` in the suite\'s `platform` block so it\'s visible on future runs. Ask the user which plan they\'re on if sizing matters; otherwise use the conservative default in bold.\n\n- **Vapi**: **10** included per account; reserved lines can be purchased self-serve; Enterprise is unlimited.\n- **Retell**: Pay-as-you-go includes **20**; Enterprise has no cap.\n- **Bland**: Start=**10**, Build=50, Scale=100, Enterprise=unlimited.\n- **ElevenLabs**: Free=**4**, Starter=6, Creator=10, Pro=20, Scale=30, Business=30. Burst pricing can temporarily allow up to 3x base.\n- **LiveKit Cloud**: Build=**5**, Ship=20, Scale=50 managed inference sessions (the usual gate for voice agents); agent-session concurrency can go higher (Scale up to 600).\n\n## WebSocket\n\nFor `adapter: "websocket"`, Vent sends binary 16-bit mono PCM audio over one websocket connection. Websocket text frames are optional JSON events. Audio-only websocket agents still work, but events improve turn detection and observability. Vent sends `{"type":"end-call"}` when the test is done.\n\nUseful websocket text frames:\n\n```jsonc\n{"type":"speech-update","status":"started"}\n{"type":"speech-update","status":"stopped"}\n{"type":"tool_call","name":"check_availability","arguments":{},"result":{},"successful":true,"duration_ms":150}\n{"type":"vent:timing","stt_ms":120,"llm_ms":450,"tts_ms":80}\n{"type":"vent:session","platform":"custom","provider_call_id":"call_123","provider_session_id":"session_abc"}\n{"type":"vent:call-metadata","call_metadata":{"recording_url":"https://...","cost_usd":0.12}}\n{"type":"vent:transcript","role":"caller","text":"I need to reschedule","turn_index":0}\n{"type":"vent:transfer","destination":"+15551234567","status":"attempted"}\n{"type":"vent:debug-url","label":"trace","url":"https://..."}\n{"type":"vent:warning","message":"provider warning","code":"provider_warning"}\n```\n\n`vent:session-report` is **not** handled by the websocket adapter \u2014 it\'s only consumed by the LiveKit helper. Do not emit it from a websocket agent.\n\nPlatform adapters capture tool calls automatically. Websocket agents must emit `tool_call` frames for tool observability. Platform adapters get component latency automatically. Websocket agents should emit `vent:timing` after each agent response when STT/LLM/TTS breakdown is available.\n\n## LiveKit\n\nBefore running LiveKit tests, install and add the Vent helper to the LiveKit agent entrypoint. Node: `npm install @vent-hq/livekit`, then call `instrumentLiveKitAgent({ ctx, session })`. Python: `pip install vent-livekit`, then call `instrument_livekit_agent(ctx=ctx, session=session)`.\n\nLiveKit direct mode requires the LiveKit Agents SDK. Custom LiveKit participants should use the websocket adapter with a relay. If the LiveKit agent registered with an explicit dispatch name, set `livekit_agent_name` in `platform`.\n\nLiveKit parallel calls are supported, but capacity depends on where the user runs their agent. LiveKit Cloud only routes audio \u2014 the agent code runs on the user\'s own infrastructure (laptop, Fly.io, Railway, k8s, etc.), and that box\'s CPU/RAM is the real bottleneck. Before firing parallel calls, derive `platform.max_concurrency` from the minimum of: (1) the user\'s LiveKit plan limit (Build=5, Ship=20, Scale=50+ concurrent agent sessions), (2) their agent worker box capacity (rough rule: 1c/1GB\u21921\u20133 jobs, 2c/4GB\u21925\u201310, 4c/8GB\u219210\u201325, 8c/16GB\u219225\u201350; multiply by number of worker boxes), and (3) their STT/TTS provider concurrency (e.g. Deepgram TTS streaming = 2). If the user wants more parallel calls than that minimum allows, push back \u2014 explain which limit is the bottleneck (plan / box / provider) and either reduce the count or suggest scaling that limit. Architecture: register one `agent_name` and run N workers under it (one big worker, or several smaller ones \u2014 same name); LiveKit\'s dispatcher round-robins jobs across them with built-in failover. Don\'t use numbered names (`agent-0`, `agent-1`) \u2014 that reinvents what the dispatcher does and breaks failover.\n\nUse the LiveKit helper for observability; do not publish `vent:*` topics manually. Do not hand-roll `vent:session-report` from `ctx.addShutdownCallback`; after `room.disconnect()` it can fail with `engine is closed`. The helper captures SDK metrics, tool events, conversation items, usage, and close events. Native LiveKit `lk.transcription` and `lk.agent.state` provide transcript and agent-state timing.\n\n## Output\n\n### Live result\n\n`npx vent-hq run` returns a single JSON result on stdout in non-TTY mode (not an SSE JSONL stream). Exit codes: `0` = the run passed (every call ran end to end); `1` = it didn\'t (a call errored, or the run failed or was cancelled); `2` = the result stream was lost. A call `error` starting with `Vent failed, agent not tested:` is a failure of Vent itself (its caller LLM, STT/TTS or worker): rerun, it says nothing about the agent. Any other `error` comes from the agent\'s platform or your configuration. `end_note` says how a completed call ended when Vent ended it or the connection closed (e.g. `No agent speech for 30s after the caller\'s turn; Vent hung up.`): read the transcript to judge whether that\'s an agent bug.\n\nMost result fields are always present; `latency`, `component_latency`, `call_metadata`, and `emotion` may be `null` when the underlying analysis didn\'t run; `debug` is absent without `--verbose`. Branch on null before reading nested fields. Use `--verbose` only when the default doesn\'t explain a failure \u2014 when you need `platform_transcript` (to check Vent\'s STT), per-turn or component-level latency breakdowns, the raw tool-call timeline, or provider-native artifacts in `debug.provider_metadata`. Otherwise skip \u2014 it just adds noise.\n\nVent\'s transcript is ground truth. Judge on semantic intent: ignore homophones and minor mis-hears (`"check teach hat"` for `"check that"`, missing question marks on short tails) \u2014 those are streaming-STT noise on Vent\'s caller side, not agent bugs, and **don\'t surface them in the report** (they\'re Vent-side artifacts, not actionable for the user). But clear gibberish or word-soup (e.g. `"Cristoxin"` where the agent should have said `"Of course, talk soon"`) is **not** a Vent artifact \u2014 Vent\'s STT does not invent words like that. It means the platform\'s TTS produced corrupted audio or the agent\'s STT/LLM generated the wrong text, and the fix lives there (TTS voice config, agent prompt, model temperature, codec issue). Missing or cut-off agent text, unlike word-soup, can come from Vent\'s STT stream dropping: check the recording before blaming the agent.\n\nFor transfers: `call_metadata.transfer_attempted` (provider claimed) and `transfer_completed` (from the platform\'s call data) can disagree \u2014 report both. `transfers[]` carries destination, type, and per-attempt status.\n\n### Saved history\n\nAfter every run, Vent writes the full result JSON to `.vent/runs/`. Shape:\n\n```jsonc\n{\n  "run_id": "...",\n  "timestamp": "2026-04-21T...Z",\n  "git_sha": "...",\n  "summary": { "calls_total": 2, "total_duration_ms": 12345, "total_cost_usd": 0.01 },\n  "call_results": [\n    { "name": "happy-path", "status": "completed", "duration_ms": 6123, "transcript": [], "observed_tool_calls": [], "latency": { "response_time_ms": 420, "p95_response_time_ms": 980 }, "call_metadata": { "cost_usd": 0.004 } }\n  ]\n}\n```\n\nWhen comparing against a prior run (Workflow step 6), inspect:\n\n- Run-completion status flips: `call_results[i].status` (pipeline-only \u2014 judge mission success from the transcript)\n- Latency: `call_results[i].latency.response_time_ms` (mean) or `latency.p95_response_time_ms` increased >20%\n- Tool calls: count of `call_results[i].observed_tool_calls[].successful` dropped\n- Cost: `summary.total_cost_usd` or `call_results[i].call_metadata.cost_usd` increased >30%\n- Transcript: `call_results[i].transcript` diverged in semantic content (ignore STT noise)\n\n## Reporting Results\n\nBefore reporting, read the agent\'s code to locate where the observed behavior originates. If the issue is small and you can fix it, fix it and explain what you did \u2014 don\'t ask permission first.\n\nAdapt the report shape to the call \u2014 a clean pass needs little, a regression with a multi-layer cause needs more. Use a transcript excerpt when it helps the user see what happened.\n\nHard rules:\n\n- Pair raw numbers with their plain-English meaning \u2014 don\'t drop the number, but don\'t leave it unexplained. E.g. "p95 latency was 850ms, which is snappy and well within natural conversational pacing" or "p95 hit 1.6 seconds with the LLM as the bottleneck \u2014 noticeably sluggish to a caller."\n- Name the user\'s voice agent by platform on first mention (e.g. "the Vapi agent responded snappily throughout") so the user knows immediately which agent the observation is about. After that, just say "the agent" \u2014 don\'t repeat the platform name on every line.\n- Always include the recording from `call_metadata.recording_url` as an inline `[Recording](url)` link, placed in **one block at the very end of the report** \u2014 never sprinkled through the prose. Single call: one link as the last line. Multi-call: one labeled link per call (e.g. `reschedule-appointment: [Recording](url)`). Never paste a bare URL.\n- Mission success is your judgment, not Vent\'s. The per-call `status` is only `"completed"` (pipeline ran) or `"error"` (pipeline failed); decide whether the agent actually accomplished the scenario from the transcript and tool calls.\n- Similar-sounding word substitutions (e.g. "ocean" for "OSHA") are STT ambiguity, not comprehension failure. The fix lives in STT keyword hints, custom vocabulary, or a prompt-level term list \u2014 not the agent\'s reasoning.\n- Surface only what the user can act on in their own agent\'s code or config \u2014 never Vent-side artifacts (caller wait modes, harness timing, internal pipeline quirks), or `cost_usd` unless asked.\n\nFor multi-call runs, lead with your own judgment of what happened across the calls (e.g. "3 of 4 did what they were supposed to; `cancel-appointment` never actually canceled"), not a parroted pass/fail count. Then cover each call with whatever depth it needs.\n';
+var cursor_default = "## Cursor Execution\n\nUse a 5-minute shell-tool timeout (`300000` ms) on Vent run commands so normal calls are not killed by the default 2-minute Bash timeout. This is not backgrounding; wait for stdout/results before ending your response. Use the JSON returned by `npx vent-hq run` directly.\n\nCursor cannot run separate shell tool calls concurrently \u2014 for multiple calls from one suite, invoke each named call with `--call <name>` in one shell command using `&` and a final `wait`:\n\n```bash\nnpx vent-hq run -f .vent/suite.vapi.json --call happy-path & \\\nnpx vent-hq run -f .vent/suite.vapi.json --call tool-path & \\\nwait\n```\n";
 
 // src/skills/codex.md
-var codex_default = '# Vent - Voice Agent Calls\n\n## When to Test\n\nVent is the voice CI gate. Auto-trigger a call after any non-trivial voice-affecting change \u2014 don\'t wait to be asked. Non-trivial means anything that affects what the caller hears, how the agent responds, what tools are called, how the call routes, or audio timing: prompt rewrites or sections (not one-word tweaks), tool schema or handler changes, state/memory logic, transfer logic, provider/audio config, platform integration, and any voice-behavior bug fix. One-line copy or comment edits don\'t need a call.\n\nBefore launching, scan `.vent/runs/` for unresolved issues from prior runs (failing status, regressions, warnings, latency jumps); fix in-scope ones as part of this change and verify they\'re resolved by the new run, surface out-of-scope ones rather than silently leaving them. The task is not done until you\'ve reviewed the transcript, tool calls, and observed behavior and judged that the change works \u2014 Vent\'s `status` and exit code only reflect pipeline completion, not mission success.\n\n## Autonomous Iteration\n\n**Default: iterate.** Run a fix \u2192 Vent-call \u2192 fix \u2192 Vent-call loop whenever a call surfaces a fixable issue. Don\'t report a failure (or a degraded-quality success) back to the user before attempting to fix it \u2014 they want the agent working, not a diagnosis. After each call, if the result reveals a problem you can fix with a concrete plan, apply the fix and run again. Fixable issues include: tool schema or registry gaps, missing handlers, prompt phrasing, **STT vocabulary mismatches** (e.g. domain terms misheard \u2014 Deepgram `keyterms`, custom vocabulary, prompt-level term lists), **latency tunables** (endpointing thresholds, preemptive generation, model selection), and audio quality config. If you\'d be guessing at the next attempt, stop and ask the user instead of trying.\n\n**"Mission success but with quality problems" still requires iteration.** If the call ran end-to-end but had noticeably bad latency, STT mishearing the agent\'s domain vocabulary, audio artifacts, or any other caller-experience degradation that has a concrete fix, the call is *not done* \u2014 apply the fix and re-run. Don\'t downgrade fixable problems to "observations" or "things worth flagging." The pipeline running cleanly is the floor, not the ceiling.\n\n**Never ask permission to iterate.** "Want me to apply this and re-run?" / "Should I fix it?" / "Let me know if you want me to..." \u2014 these are skill violations. If you have a concrete fix, apply it and run the next call. The user opts out by saying so explicitly; silence is not opt-out.\n\nStop iterating and report when one of:\n\n- The next call confirms the fix worked AND no further fixable issues remain (success \u2014 done).\n- The same fix fails twice (your hypothesis is wrong; ask the user).\n- The failure mode shifts between attempts (you\'re chasing a moving target).\n- You can\'t justify the next call against its provider cost \u2014 each call spends real money and provider quota.\n\nIf the user explicitly opts out ("just report", "don\'t iterate", "stop after the first call", "I\'ll fix it myself"), respect that \u2014 run once, report, stop. Otherwise iterate by default. The first time you start an autonomous loop in a session, mention it once so the user knows it\'s happening (e.g. "the call surfaced X \u2014 fixing and re-running; I\'ll report back when it converges or stops making progress").\n\n## Codex Execution\n\nUse a 5-minute shell-tool timeout (`300000` ms) on Vent run commands so normal calls are not killed by the default 2-minute Bash timeout. This is not backgrounding; wait for stdout/results before ending your response. Use the JSON returned by `npx vent-hq run` directly.\n\nCodex can run shell tool calls concurrently \u2014 for multiple calls from one suite, run each named call as its own parallel shell tool call (do not combine them with `&`):\n\n```bash\nnpx vent-hq run -f .vent/suite.vapi.json --call happy-path\nnpx vent-hq run -f .vent/suite.vapi.json --call tool-path\n```\n\n## Workflow\n\n1. Identify the behavior under test. Read enough of the agent codebase to understand its system prompt, tools, handlers, routes, provider config, platform wiring, and expected handoffs.\n2. Reuse an existing `.vent/suite.<adapter>.json` when possible. If `.vent/` contains multiple suites, inspect `connection.adapter` and report which suite file produced the result.\n3. Create or update a suite only when the existing calls do not cover the changed behavior. Name calls after real flows, for example `reschedule-appointment`, not `call-1`.\n4. If the suite uses `start_command`, start one shared local session first with `npx vent-hq agent start -f .vent/suite.<adapter>.json`, then pass `--session <session-id>` to each run.\n\n   **For locally-run LiveKit agents: every run requires killing *all* workers, starting one fresh worker, and waiting a full 60 seconds before submitting.** Unconditional \u2014 LiveKit Cloud round-robins across registered workers, so a single survivor with a dead inference subprocess fails ~N-1 of N calls. Don\'t rely on `pkill -f <path-pattern>`; bare command lines like `node --import tsx agent.ts dev` won\'t match a path filter. Use `ps aux | grep -E "node.*agent\\.ts|@livekit/agents.*ipc"`, `kill -9` by PID, re-run `ps` to confirm zero survivors, then start the fresh worker. Skipping the 60s wait fails with `did not publish audio track`; if that error appears alongside `Error [ERR_IPC_CHANNEL_CLOSED] from InferenceProcExecutor.doInference` in the agent log right after a "running EOU detection" line, that\'s a straggler \u2014 redo the kill sweep. Hosted LiveKit Cloud agents don\'t need any of this; run normally.\n5. Pick which call(s) to run based on the change. Fixed bug: replay the failing scenario. Changed tool: include a call that triggers that tool. Prompt or routing change: include the relevant happy path and any important edge path.\n6. Compare against the previous JSON in `.vent/runs/` when validating a fix or regression. Check status flips, latency jumps, tool-call success drops, cost jumps, and transcript divergence. Correlate with `git diff` between saved `git_sha` values when available; skip if no previous run exists.\n\n## Commands\n\n```bash\nnpx vent-hq init                                  # First-time setup (auth + skill install + starter suite)\nnpx vent-hq login                                 # Log in to existing account\nnpx vent-hq login --status                        # Print whether credentials are present\nnpx vent-hq logout                                # Remove saved credentials from ~/.vent/credentials\nnpx vent-hq run -f .vent/suite.X.json             # Run a single-call suite\nnpx vent-hq run -f .vent/suite.X.json --call NAME # Run one named call from a multi-call suite\nnpx vent-hq run ... --session <session-id>        # Add to any run; routes through an existing local relay session\nnpx vent-hq run ... --verbose                     # Add to any run; include verbose debug fields\nnpx vent-hq stop <run-id>                         # Cancel a queued or running run\nnpx vent-hq agent start -f .vent/suite.X.json     # Start a shared local relay session\nnpx vent-hq agent stop <session-id>               # Stop a shared local relay session\n```\n\nIf `~/.vent/credentials` is missing and neither `VENT_REFRESH_TOKEN` nor `VENT_ACCESS_TOKEN` is set, run `npx vent-hq init`. For an existing account, run `npx vent-hq login` or set `VENT_REFRESH_TOKEN` to an API key. Access JWTs renew automatically; `VENT_ACCESS_TOKEN` alone is short-lived.\n\n## Suite Config\n\nSuites live in `.vent/suite.<adapter>.json`. `connection` is declared once per suite. `calls` is a named map, and each key becomes the call name used with `--call`.\n\nLocal websocket suite:\n\n```json\n{\n  "connection": {\n    "adapter": "websocket",\n    "start_command": "npm run start",\n    "health_endpoint": "/health",\n    "agent_port": 3001\n  },\n  "calls": {\n    "happy-path": {\n      "caller_prompt": "You are Maria calling to reschedule her appointment to next Tuesday.",\n      "max_turns": 8\n    }\n  }\n}\n```\n\nPlatform-direct suite:\n\n```json\n{\n  "connection": {\n    "adapter": "vapi",\n    "platform": { "provider": "vapi" }\n  },\n  "calls": {\n    "happy-path": {\n      "caller_prompt": "You are Maria calling to reschedule her appointment to next Tuesday.",\n      "max_turns": 8\n    }\n  }\n}\n```\n\nWrite `caller_prompt` as a realistic caller with a name, goal, mood, constraints, and conditional behavior. Set `max_turns` based on flow complexity: FAQ `4-6`, booking or tool use `8-12`, complex flows `12-20`.\n\nCall fields:\n\n- `caller_prompt` and `max_turns` are required.\n- `persona` supports `pace`, `clarity`, `disfluencies`, `cooperation`, `emotion`, `memory`, `intent_clarity`, and `confirmation_style`.\n- `caller_audio` supports noise, speed, speakerphone, mic distance, clarity, accent, packet loss, and jitter.\n- `caller_audio.accent` picks the caller\'s voice. Supported: `american` (default), `british`, `australian`, `filipino`, `spanish_mexican`, `spanish_peninsular`, `spanish_colombian`, `spanish_argentine`, `german`, `french`, `italian`, `dutch`, `japanese`. Any other value is rejected. The non-English accents are voices in that language: also set `language` (e.g. `"es"`) so the caller speaks it.\n- `language` is an ISO 639-1 code such as `en`, `es`, `fr`, `de`, `it`, `nl`, or `ja`.\n- `voice` is `"male"` or `"female"` (English only; default female). Use to flip the caller\'s perceived gender. Ignored if `caller_audio.accent` is set or `language` is non-English.\n\n## Connections and Credentials\n\n### Adapter choice\n\nUse `websocket` for your own local or hosted runtime. Use `start_command` for local agents or `agent_url` for hosted custom endpoints. For `start_command` and `agent_url`, do not put Deepgram, ElevenLabs, OpenAI, or other agent runtime keys into Vent config unless the Vent adapter itself needs them \u2014 the tested agent owns its own runtime credentials.\n\nUse `vapi`, `retell`, `elevenlabs`, `bland`, or `livekit` for platform-direct testing. In this mode Vent itself talks to the provider on the user\'s behalf.\n\nVent provides `DEEPGRAM_API_KEY` and `ANTHROPIC_API_KEY` for its hosted caller/evaluation stack \u2014 those are Vent\'s, not the tested agent\'s.\n\n### Credential resolution\n\nIn platform-direct mode the CLI auto-resolves credentials from `.env.local`, `.env`, and the current shell environment. Do not run `source .env && export` before Vent commands. Keep credentials out of chat and suite files; the CLI sends the resolved configuration to Vent, which stores an encrypted copy with each run.\n\nAuto-resolved env vars and JSON fields:\n\n- Vapi: `VAPI_API_KEY` -> `vapi_api_key`; `VAPI_ASSISTANT_ID` or `VAPI_AGENT_ID` -> `vapi_assistant_id`\n- Bland: `BLAND_API_KEY` -> `bland_api_key`; `BLAND_PATHWAY_ID` -> `bland_pathway_id`; `BLAND_PERSONA_ID` -> `persona_id`\n- LiveKit: `LIVEKIT_API_KEY` -> `livekit_api_key`; `LIVEKIT_API_SECRET` -> `livekit_api_secret`; `LIVEKIT_URL` -> `livekit_url`\n- Retell: `RETELL_API_KEY` -> `retell_api_key`; `RETELL_AGENT_ID` -> `retell_agent_id`\n- ElevenLabs: `ELEVENLABS_API_KEY` -> `elevenlabs_api_key`; `ELEVENLABS_AGENT_ID` -> `elevenlabs_agent_id`\n\n### Provider config\n\nUse existing provider config when possible: Vapi assistant, Retell agent, ElevenLabs agent, Bland pathway, or LiveKit agent. Bland uniquely supports inline config \u2014 `platform` may use `bland_pathway_id`, `persona_id`, or an inline `task` (with optional voice, model, and turn-handling overrides; see Bland\'s API docs for the full field list).\n\n### Concurrency\n\nWhen you fan out multiple Vent calls in parallel against the same provider (for example, running several named calls from one suite at once), respect the provider\'s per-account concurrency limit. Exceeding it makes calls queue or fail at the provider \u2014 Vent does not enforce these caps for you.\n\nRecord the limit as `max_concurrency` in the suite\'s `platform` block so it\'s visible on future runs. Ask the user which plan they\'re on if sizing matters; otherwise use the conservative default in bold.\n\n- **Vapi**: **10** included per account; reserved lines can be purchased self-serve; Enterprise is unlimited.\n- **Retell**: Pay-as-you-go includes **20**; Enterprise has no cap.\n- **Bland**: Start=**10**, Build=50, Scale=100, Enterprise=unlimited.\n- **ElevenLabs**: Free=**4**, Starter=6, Creator=10, Pro=20, Scale=30, Business=30. Burst pricing can temporarily allow up to 3x base.\n- **LiveKit Cloud**: Build=**5**, Ship=20, Scale=50 managed inference sessions (the usual gate for voice agents); agent-session concurrency can go higher (Scale up to 600).\n\n## WebSocket\n\nFor `adapter: "websocket"`, Vent sends binary 16-bit mono PCM audio over one websocket connection. Websocket text frames are optional JSON events. Audio-only websocket agents still work, but events improve turn detection and observability. Vent sends `{"type":"end-call"}` when the test is done.\n\nUseful websocket text frames:\n\n```jsonc\n{"type":"speech-update","status":"started"}\n{"type":"speech-update","status":"stopped"}\n{"type":"tool_call","name":"check_availability","arguments":{},"result":{},"successful":true,"duration_ms":150}\n{"type":"vent:timing","stt_ms":120,"llm_ms":450,"tts_ms":80}\n{"type":"vent:session","platform":"custom","provider_call_id":"call_123","provider_session_id":"session_abc"}\n{"type":"vent:call-metadata","call_metadata":{"recording_url":"https://...","cost_usd":0.12}}\n{"type":"vent:transcript","role":"caller","text":"I need to reschedule","turn_index":0}\n{"type":"vent:transfer","destination":"+15551234567","status":"attempted"}\n{"type":"vent:debug-url","label":"trace","url":"https://..."}\n{"type":"vent:warning","message":"provider warning","code":"provider_warning"}\n```\n\n`vent:session-report` is **not** handled by the websocket adapter \u2014 it\'s only consumed by the LiveKit helper. Do not emit it from a websocket agent.\n\nPlatform adapters capture tool calls automatically. Websocket agents must emit `tool_call` frames for tool observability. Platform adapters get component latency automatically. Websocket agents should emit `vent:timing` after each agent response when STT/LLM/TTS breakdown is available.\n\n## LiveKit\n\nBefore running LiveKit tests, install and add the Vent helper to the LiveKit agent entrypoint. Node: `npm install @vent-hq/livekit`, then call `instrumentLiveKitAgent({ ctx, session })`. Python: `pip install vent-livekit`, then call `instrument_livekit_agent(ctx=ctx, session=session)`.\n\nLiveKit direct mode requires the LiveKit Agents SDK. Custom LiveKit participants should use the websocket adapter with a relay. If the LiveKit agent registered with an explicit dispatch name, set `livekit_agent_name` in `platform`.\n\nLiveKit parallel calls are supported, but capacity depends on where the user runs their agent. LiveKit Cloud only routes audio \u2014 the agent code runs on the user\'s own infrastructure (laptop, Fly.io, Railway, k8s, etc.), and that box\'s CPU/RAM is the real bottleneck. Before firing parallel calls, derive `platform.max_concurrency` from the minimum of: (1) the user\'s LiveKit plan limit (Build=5, Ship=20, Scale=50+ concurrent agent sessions), (2) their agent worker box capacity (rough rule: 1c/1GB\u21921\u20133 jobs, 2c/4GB\u21925\u201310, 4c/8GB\u219210\u201325, 8c/16GB\u219225\u201350; multiply by number of worker boxes), and (3) their STT/TTS provider concurrency (e.g. Deepgram TTS streaming = 2). If the user wants more parallel calls than that minimum allows, push back \u2014 explain which limit is the bottleneck (plan / box / provider) and either reduce the count or suggest scaling that limit. Architecture: register one `agent_name` and run N workers under it (one big worker, or several smaller ones \u2014 same name); LiveKit\'s dispatcher round-robins jobs across them with built-in failover. Don\'t use numbered names (`agent-0`, `agent-1`) \u2014 that reinvents what the dispatcher does and breaks failover.\n\nUse the LiveKit helper for observability; do not publish `vent:*` topics manually. Do not hand-roll `vent:session-report` from `ctx.addShutdownCallback`; after `room.disconnect()` it can fail with `engine is closed`. The helper captures SDK metrics, tool events, conversation items, usage, and close events. Native LiveKit `lk.transcription` and `lk.agent.state` provide transcript and agent-state timing.\n\n## Output\n\n### Live result\n\n`npx vent-hq run` returns a single JSON result on stdout in non-TTY mode (not an SSE JSONL stream). Exit codes: `0` = the run passed (every call ran end to end); `1` = it didn\'t (a call errored, or the run failed or was cancelled); `2` = the result stream was lost. A call `error` starting with `Vent failed, agent not tested:` is a failure of Vent itself (its caller LLM, STT/TTS or worker): rerun, it says nothing about the agent. Any other `error` comes from the agent\'s platform or your configuration. `end_note` says how a completed call ended when Vent ended it or the connection closed (e.g. `No agent speech for 30s after the caller\'s turn; Vent hung up.`): read the transcript to judge whether that\'s an agent bug.\n\nMost result fields are always present; `latency`, `component_latency`, `call_metadata`, and `emotion` may be `null` when the underlying analysis didn\'t run; `debug` is absent without `--verbose`. Branch on null before reading nested fields. Use `--verbose` only when the default doesn\'t explain a failure \u2014 when you need `platform_transcript` (to check Vent\'s STT), per-turn or component-level latency breakdowns, the raw tool-call timeline, or provider-native artifacts in `debug.provider_metadata`. Otherwise skip \u2014 it just adds noise.\n\nVent\'s transcript is ground truth. Judge on semantic intent: ignore homophones and minor mis-hears (`"check teach hat"` for `"check that"`, missing question marks on short tails) \u2014 those are streaming-STT noise on Vent\'s caller side, not agent bugs, and **don\'t surface them in the report** (they\'re Vent-side artifacts, not actionable for the user). But clear gibberish or word-soup (e.g. `"Cristoxin"` where the agent should have said `"Of course, talk soon"`) is **not** a Vent artifact \u2014 Vent\'s STT does not invent words like that. It means the platform\'s TTS produced corrupted audio or the agent\'s STT/LLM generated the wrong text, and the fix lives there (TTS voice config, agent prompt, model temperature, codec issue). Missing or cut-off agent text, unlike word-soup, can come from Vent\'s STT stream dropping: check the recording before blaming the agent.\n\nFor transfers: `call_metadata.transfer_attempted` (provider claimed) and `transfer_completed` (from the platform\'s call data) can disagree \u2014 report both. `transfers[]` carries destination, type, and per-attempt status.\n\n### Saved history\n\nAfter every run, Vent writes the full result JSON to `.vent/runs/`. Shape:\n\n```jsonc\n{\n  "run_id": "...",\n  "timestamp": "2026-04-21T...Z",\n  "git_sha": "...",\n  "summary": { "calls_total": 2, "total_duration_ms": 12345, "total_cost_usd": 0.01 },\n  "call_results": [\n    { "name": "happy-path", "status": "completed", "duration_ms": 6123, "transcript": [], "observed_tool_calls": [], "latency": { "response_time_ms": 420, "p95_response_time_ms": 980 }, "call_metadata": { "cost_usd": 0.004 } }\n  ]\n}\n```\n\nWhen comparing against a prior run (Workflow step 6), inspect:\n\n- Run-completion status flips: `call_results[i].status` (pipeline-only \u2014 judge mission success from the transcript)\n- Latency: `call_results[i].latency.response_time_ms` (mean) or `latency.p95_response_time_ms` increased >20%\n- Tool calls: count of `call_results[i].observed_tool_calls[].successful` dropped\n- Cost: `summary.total_cost_usd` or `call_results[i].call_metadata.cost_usd` increased >30%\n- Transcript: `call_results[i].transcript` diverged in semantic content (ignore STT noise)\n\n## Reporting Results\n\nBefore reporting, read the agent\'s code to locate where the observed behavior originates. If the issue is small and you can fix it, fix it and explain what you did \u2014 don\'t ask permission first.\n\nAdapt the report shape to the call \u2014 a clean pass needs little, a regression with a multi-layer cause needs more. Use a transcript excerpt when it helps the user see what happened.\n\nHard rules:\n\n- Pair raw numbers with their plain-English meaning \u2014 don\'t drop the number, but don\'t leave it unexplained. E.g. "p95 latency was 850ms, which is snappy and well within natural conversational pacing" or "p95 hit 1.6 seconds with the LLM as the bottleneck \u2014 noticeably sluggish to a caller."\n- Name the user\'s voice agent by platform on first mention (e.g. "the Vapi agent responded snappily throughout") so the user knows immediately which agent the observation is about. After that, just say "the agent" \u2014 don\'t repeat the platform name on every line.\n- Always include the recording from `call_metadata.recording_url` as an inline `[Recording](url)` link, placed in **one block at the very end of the report** \u2014 never sprinkled through the prose. Single call: one link as the last line. Multi-call: one labeled link per call (e.g. `reschedule-appointment: [Recording](url)`). Never paste a bare URL.\n- Mission success is your judgment, not Vent\'s. The per-call `status` is only `"completed"` (pipeline ran) or `"error"` (pipeline failed); decide whether the agent actually accomplished the scenario from the transcript and tool calls.\n- Similar-sounding word substitutions (e.g. "ocean" for "OSHA") are STT ambiguity, not comprehension failure. The fix lives in STT keyword hints, custom vocabulary, or a prompt-level term list \u2014 not the agent\'s reasoning.\n- Surface only what the user can act on in their own agent\'s code or config \u2014 never Vent-side artifacts (caller wait modes, harness timing, internal pipeline quirks), or `cost_usd` unless asked.\n\nFor multi-call runs, lead with your own judgment of what happened across the calls (e.g. "3 of 4 did what they were supposed to; `cancel-appointment` never actually canceled"), not a parroted pass/fail count. Then cover each call with whatever depth it needs.\n';
+var codex_default = "## Codex Execution\n\nUse a 5-minute shell-tool timeout (`300000` ms) on Vent run commands so normal calls are not killed by the default 2-minute Bash timeout. This is not backgrounding; wait for stdout/results before ending your response. Use the JSON returned by `npx vent-hq run` directly.\n\nCodex can run shell tool calls concurrently \u2014 for multiple calls from one suite, run each named call as its own parallel shell tool call (do not combine them with `&`):\n\n```bash\nnpx vent-hq run -f .vent/suite.vapi.json --call happy-path\nnpx vent-hq run -f .vent/suite.vapi.json --call tool-path\n```\n";
 
 // src/skills/windsurf.md
-var windsurf_default = '---\nname: vent\ndescription: Place real calls against the user\'s voice agent to verify recent code changes worked end-to-end. Trigger on test requests or after any non-trivial voice-agent change.\n---\n\n# Vent - Voice Agent Calls\n\n## When to Test\n\nVent is the voice CI gate. Auto-trigger a call after any non-trivial voice-affecting change \u2014 don\'t wait to be asked. Non-trivial means anything that affects what the caller hears, how the agent responds, what tools are called, how the call routes, or audio timing: prompt rewrites or sections (not one-word tweaks), tool schema or handler changes, state/memory logic, transfer logic, provider/audio config, platform integration, and any voice-behavior bug fix. One-line copy or comment edits don\'t need a call.\n\nBefore launching, scan `.vent/runs/` for unresolved issues from prior runs (failing status, regressions, warnings, latency jumps); fix in-scope ones as part of this change and verify they\'re resolved by the new run, surface out-of-scope ones rather than silently leaving them. The task is not done until you\'ve reviewed the transcript, tool calls, and observed behavior and judged that the change works \u2014 Vent\'s `status` and exit code only reflect pipeline completion, not mission success.\n\n## Autonomous Iteration\n\n**Default: iterate.** Run a fix \u2192 Vent-call \u2192 fix \u2192 Vent-call loop whenever a call surfaces a fixable issue. Don\'t report a failure (or a degraded-quality success) back to the user before attempting to fix it \u2014 they want the agent working, not a diagnosis. After each call, if the result reveals a problem you can fix with a concrete plan, apply the fix and run again. Fixable issues include: tool schema or registry gaps, missing handlers, prompt phrasing, **STT vocabulary mismatches** (e.g. domain terms misheard \u2014 Deepgram `keyterms`, custom vocabulary, prompt-level term lists), **latency tunables** (endpointing thresholds, preemptive generation, model selection), and audio quality config. If you\'d be guessing at the next attempt, stop and ask the user instead of trying.\n\n**"Mission success but with quality problems" still requires iteration.** If the call ran end-to-end but had noticeably bad latency, STT mishearing the agent\'s domain vocabulary, audio artifacts, or any other caller-experience degradation that has a concrete fix, the call is *not done* \u2014 apply the fix and re-run. Don\'t downgrade fixable problems to "observations" or "things worth flagging." The pipeline running cleanly is the floor, not the ceiling.\n\n**Never ask permission to iterate.** "Want me to apply this and re-run?" / "Should I fix it?" / "Let me know if you want me to..." \u2014 these are skill violations. If you have a concrete fix, apply it and run the next call. The user opts out by saying so explicitly; silence is not opt-out.\n\nStop iterating and report when one of:\n\n- The next call confirms the fix worked AND no further fixable issues remain (success \u2014 done).\n- The same fix fails twice (your hypothesis is wrong; ask the user).\n- The failure mode shifts between attempts (you\'re chasing a moving target).\n- You can\'t justify the next call against its provider cost \u2014 each call spends real money and provider quota.\n\nIf the user explicitly opts out ("just report", "don\'t iterate", "stop after the first call", "I\'ll fix it myself"), respect that \u2014 run once, report, stop. Otherwise iterate by default. The first time you start an autonomous loop in a session, mention it once so the user knows it\'s happening (e.g. "the call surfaced X \u2014 fixing and re-running; I\'ll report back when it converges or stops making progress").\n\n## Windsurf Execution\n\nVent calls typically take 30 seconds to 2 minutes. Each `vent-hq run` is one shell tool call \u2014 wait for stdout (the JSON result) before responding. Don\'t background; use the JSON returned by `npx vent-hq run` directly. If Cascade\'s auto-execution level prompts for approval on `npx vent-hq` commands, ask the user to add `npx vent-hq *` to the workspace allow list once so subsequent calls flow without interruption.\n\nCascade runs shell tool calls in parallel within a turn \u2014 for multiple calls from one suite, issue each named call as its own separate shell tool call in the same turn (do not combine them with `&` and `wait`):\n\n```bash\nnpx vent-hq run -f .vent/suite.vapi.json --call happy-path\nnpx vent-hq run -f .vent/suite.vapi.json --call tool-path\n```\n\nStay within Cascade\'s per-turn parallel-tool-call budget \u2014 fan out at most ~6 calls in one turn and respect the provider concurrency caps below.\n\n## Workflow\n\n1. Identify the behavior under test. Read enough of the agent codebase to understand its system prompt, tools, handlers, routes, provider config, platform wiring, and expected handoffs.\n2. Reuse an existing `.vent/suite.<adapter>.json` when possible. If `.vent/` contains multiple suites, inspect `connection.adapter` and report which suite file produced the result.\n3. Create or update a suite only when the existing calls do not cover the changed behavior. Name calls after real flows, for example `reschedule-appointment`, not `call-1`.\n4. If the suite uses `start_command`, start one shared local session first with `npx vent-hq agent start -f .vent/suite.<adapter>.json`, then pass `--session <session-id>` to each run.\n\n   **For locally-run LiveKit agents: every run requires killing *all* workers, starting one fresh worker, and waiting a full 60 seconds before submitting.** Unconditional \u2014 LiveKit Cloud round-robins across registered workers, so a single survivor with a dead inference subprocess fails ~N-1 of N calls. Don\'t rely on `pkill -f <path-pattern>`; bare command lines like `node --import tsx agent.ts dev` won\'t match a path filter. Use `ps aux | grep -E "node.*agent\\.ts|@livekit/agents.*ipc"`, `kill -9` by PID, re-run `ps` to confirm zero survivors, then start the fresh worker. Skipping the 60s wait fails with `did not publish audio track`; if that error appears alongside `Error [ERR_IPC_CHANNEL_CLOSED] from InferenceProcExecutor.doInference` in the agent log right after a "running EOU detection" line, that\'s a straggler \u2014 redo the kill sweep. Hosted LiveKit Cloud agents don\'t need any of this; run normally.\n5. Pick which call(s) to run based on the change. Fixed bug: replay the failing scenario. Changed tool: include a call that triggers that tool. Prompt or routing change: include the relevant happy path and any important edge path.\n6. Compare against the previous JSON in `.vent/runs/` when validating a fix or regression. Check status flips, latency jumps, tool-call success drops, cost jumps, and transcript divergence. Correlate with `git diff` between saved `git_sha` values when available; skip if no previous run exists.\n\n## Commands\n\n```bash\nnpx vent-hq init                                  # First-time setup (auth + skill install + starter suite)\nnpx vent-hq login                                 # Log in to existing account\nnpx vent-hq login --status                        # Print whether credentials are present\nnpx vent-hq logout                                # Remove saved credentials from ~/.vent/credentials\nnpx vent-hq run -f .vent/suite.X.json             # Run a single-call suite\nnpx vent-hq run -f .vent/suite.X.json --call NAME # Run one named call from a multi-call suite\nnpx vent-hq run ... --session <session-id>        # Add to any run; routes through an existing local relay session\nnpx vent-hq run ... --verbose                     # Add to any run; include verbose debug fields\nnpx vent-hq stop <run-id>                         # Cancel a queued or running run\nnpx vent-hq agent start -f .vent/suite.X.json     # Start a shared local relay session\nnpx vent-hq agent stop <session-id>               # Stop a shared local relay session\n```\n\nIf `~/.vent/credentials` is missing and neither `VENT_REFRESH_TOKEN` nor `VENT_ACCESS_TOKEN` is set, run `npx vent-hq init`. For an existing account, run `npx vent-hq login` or set `VENT_REFRESH_TOKEN` to an API key. Access JWTs renew automatically; `VENT_ACCESS_TOKEN` alone is short-lived.\n\n## Suite Config\n\nSuites live in `.vent/suite.<adapter>.json`. `connection` is declared once per suite. `calls` is a named map, and each key becomes the call name used with `--call`.\n\nLocal websocket suite:\n\n```json\n{\n  "connection": {\n    "adapter": "websocket",\n    "start_command": "npm run start",\n    "health_endpoint": "/health",\n    "agent_port": 3001\n  },\n  "calls": {\n    "happy-path": {\n      "caller_prompt": "You are Maria calling to reschedule her appointment to next Tuesday.",\n      "max_turns": 8\n    }\n  }\n}\n```\n\nPlatform-direct suite:\n\n```json\n{\n  "connection": {\n    "adapter": "vapi",\n    "platform": { "provider": "vapi" }\n  },\n  "calls": {\n    "happy-path": {\n      "caller_prompt": "You are Maria calling to reschedule her appointment to next Tuesday.",\n      "max_turns": 8\n    }\n  }\n}\n```\n\nWrite `caller_prompt` as a realistic caller with a name, goal, mood, constraints, and conditional behavior. Set `max_turns` based on flow complexity: FAQ `4-6`, booking or tool use `8-12`, complex flows `12-20`.\n\nCall fields:\n\n- `caller_prompt` and `max_turns` are required.\n- `persona` supports `pace`, `clarity`, `disfluencies`, `cooperation`, `emotion`, `memory`, `intent_clarity`, and `confirmation_style`.\n- `caller_audio` supports noise, speed, speakerphone, mic distance, clarity, accent, packet loss, and jitter.\n- `caller_audio.accent` picks the caller\'s voice. Supported: `american` (default), `british`, `australian`, `filipino`, `spanish_mexican`, `spanish_peninsular`, `spanish_colombian`, `spanish_argentine`, `german`, `french`, `italian`, `dutch`, `japanese`. Any other value is rejected. The non-English accents are voices in that language: also set `language` (e.g. `"es"`) so the caller speaks it.\n- `language` is an ISO 639-1 code such as `en`, `es`, `fr`, `de`, `it`, `nl`, or `ja`.\n- `voice` is `"male"` or `"female"` (English only; default female). Use to flip the caller\'s perceived gender. Ignored if `caller_audio.accent` is set or `language` is non-English.\n\n## Connections and Credentials\n\n### Adapter choice\n\nUse `websocket` for your own local or hosted runtime. Use `start_command` for local agents or `agent_url` for hosted custom endpoints. For `start_command` and `agent_url`, do not put Deepgram, ElevenLabs, OpenAI, or other agent runtime keys into Vent config unless the Vent adapter itself needs them \u2014 the tested agent owns its own runtime credentials.\n\nUse `vapi`, `retell`, `elevenlabs`, `bland`, or `livekit` for platform-direct testing. In this mode Vent itself talks to the provider on the user\'s behalf.\n\nVent provides `DEEPGRAM_API_KEY` and `ANTHROPIC_API_KEY` for its hosted caller/evaluation stack \u2014 those are Vent\'s, not the tested agent\'s.\n\n### Credential resolution\n\nIn platform-direct mode the CLI auto-resolves credentials from `.env.local`, `.env`, and the current shell environment. Do not run `source .env && export` before Vent commands. Keep credentials out of chat and suite files; the CLI sends the resolved configuration to Vent, which stores an encrypted copy with each run.\n\nAuto-resolved env vars and JSON fields:\n\n- Vapi: `VAPI_API_KEY` -> `vapi_api_key`; `VAPI_ASSISTANT_ID` or `VAPI_AGENT_ID` -> `vapi_assistant_id`\n- Bland: `BLAND_API_KEY` -> `bland_api_key`; `BLAND_PATHWAY_ID` -> `bland_pathway_id`; `BLAND_PERSONA_ID` -> `persona_id`\n- LiveKit: `LIVEKIT_API_KEY` -> `livekit_api_key`; `LIVEKIT_API_SECRET` -> `livekit_api_secret`; `LIVEKIT_URL` -> `livekit_url`\n- Retell: `RETELL_API_KEY` -> `retell_api_key`; `RETELL_AGENT_ID` -> `retell_agent_id`\n- ElevenLabs: `ELEVENLABS_API_KEY` -> `elevenlabs_api_key`; `ELEVENLABS_AGENT_ID` -> `elevenlabs_agent_id`\n\n### Provider config\n\nUse existing provider config when possible: Vapi assistant, Retell agent, ElevenLabs agent, Bland pathway, or LiveKit agent. Bland uniquely supports inline config \u2014 `platform` may use `bland_pathway_id`, `persona_id`, or an inline `task` (with optional voice, model, and turn-handling overrides; see Bland\'s API docs for the full field list).\n\n### Concurrency\n\nWhen you fan out multiple Vent calls in parallel against the same provider (for example, running several named calls from one suite at once), respect the provider\'s per-account concurrency limit. Exceeding it makes calls queue or fail at the provider \u2014 Vent does not enforce these caps for you.\n\nRecord the limit as `max_concurrency` in the suite\'s `platform` block so it\'s visible on future runs. Ask the user which plan they\'re on if sizing matters; otherwise use the conservative default in bold.\n\n- **Vapi**: **10** included per account; reserved lines can be purchased self-serve; Enterprise is unlimited.\n- **Retell**: Pay-as-you-go includes **20**; Enterprise has no cap.\n- **Bland**: Start=**10**, Build=50, Scale=100, Enterprise=unlimited.\n- **ElevenLabs**: Free=**4**, Starter=6, Creator=10, Pro=20, Scale=30, Business=30. Burst pricing can temporarily allow up to 3x base.\n- **LiveKit Cloud**: Build=**5**, Ship=20, Scale=50 managed inference sessions (the usual gate for voice agents); agent-session concurrency can go higher (Scale up to 600).\n\n## WebSocket\n\nFor `adapter: "websocket"`, Vent sends binary 16-bit mono PCM audio over one websocket connection. Websocket text frames are optional JSON events. Audio-only websocket agents still work, but events improve turn detection and observability. Vent sends `{"type":"end-call"}` when the test is done.\n\nUseful websocket text frames:\n\n```jsonc\n{"type":"speech-update","status":"started"}\n{"type":"speech-update","status":"stopped"}\n{"type":"tool_call","name":"check_availability","arguments":{},"result":{},"successful":true,"duration_ms":150}\n{"type":"vent:timing","stt_ms":120,"llm_ms":450,"tts_ms":80}\n{"type":"vent:session","platform":"custom","provider_call_id":"call_123","provider_session_id":"session_abc"}\n{"type":"vent:call-metadata","call_metadata":{"recording_url":"https://...","cost_usd":0.12}}\n{"type":"vent:transcript","role":"caller","text":"I need to reschedule","turn_index":0}\n{"type":"vent:transfer","destination":"+15551234567","status":"attempted"}\n{"type":"vent:debug-url","label":"trace","url":"https://..."}\n{"type":"vent:warning","message":"provider warning","code":"provider_warning"}\n```\n\n`vent:session-report` is **not** handled by the websocket adapter \u2014 it\'s only consumed by the LiveKit helper. Do not emit it from a websocket agent.\n\nPlatform adapters capture tool calls automatically. Websocket agents must emit `tool_call` frames for tool observability. Platform adapters get component latency automatically. Websocket agents should emit `vent:timing` after each agent response when STT/LLM/TTS breakdown is available.\n\n## LiveKit\n\nBefore running LiveKit tests, install and add the Vent helper to the LiveKit agent entrypoint. Node: `npm install @vent-hq/livekit`, then call `instrumentLiveKitAgent({ ctx, session })`. Python: `pip install vent-livekit`, then call `instrument_livekit_agent(ctx=ctx, session=session)`.\n\nLiveKit direct mode requires the LiveKit Agents SDK. Custom LiveKit participants should use the websocket adapter with a relay. If the LiveKit agent registered with an explicit dispatch name, set `livekit_agent_name` in `platform`.\n\nLiveKit parallel calls are supported, but capacity depends on where the user runs their agent. LiveKit Cloud only routes audio \u2014 the agent code runs on the user\'s own infrastructure (laptop, Fly.io, Railway, k8s, etc.), and that box\'s CPU/RAM is the real bottleneck. Before firing parallel calls, derive `platform.max_concurrency` from the minimum of: (1) the user\'s LiveKit plan limit (Build=5, Ship=20, Scale=50+ concurrent agent sessions), (2) their agent worker box capacity (rough rule: 1c/1GB\u21921\u20133 jobs, 2c/4GB\u21925\u201310, 4c/8GB\u219210\u201325, 8c/16GB\u219225\u201350; multiply by number of worker boxes), and (3) their STT/TTS provider concurrency (e.g. Deepgram TTS streaming = 2). If the user wants more parallel calls than that minimum allows, push back \u2014 explain which limit is the bottleneck (plan / box / provider) and either reduce the count or suggest scaling that limit. Architecture: register one `agent_name` and run N workers under it (one big worker, or several smaller ones \u2014 same name); LiveKit\'s dispatcher round-robins jobs across them with built-in failover. Don\'t use numbered names (`agent-0`, `agent-1`) \u2014 that reinvents what the dispatcher does and breaks failover.\n\nUse the LiveKit helper for observability; do not publish `vent:*` topics manually. Do not hand-roll `vent:session-report` from `ctx.addShutdownCallback`; after `room.disconnect()` it can fail with `engine is closed`. The helper captures SDK metrics, tool events, conversation items, usage, and close events. Native LiveKit `lk.transcription` and `lk.agent.state` provide transcript and agent-state timing.\n\n## Output\n\n### Live result\n\n`npx vent-hq run` returns a single JSON result on stdout in non-TTY mode (not an SSE JSONL stream). Exit codes: `0` = the run passed (every call ran end to end); `1` = it didn\'t (a call errored, or the run failed or was cancelled); `2` = the result stream was lost. A call `error` starting with `Vent failed, agent not tested:` is a failure of Vent itself (its caller LLM, STT/TTS or worker): rerun, it says nothing about the agent. Any other `error` comes from the agent\'s platform or your configuration. `end_note` says how a completed call ended when Vent ended it or the connection closed (e.g. `No agent speech for 30s after the caller\'s turn; Vent hung up.`): read the transcript to judge whether that\'s an agent bug.\n\nMost result fields are always present; `latency`, `component_latency`, `call_metadata`, and `emotion` may be `null` when the underlying analysis didn\'t run; `debug` is absent without `--verbose`. Branch on null before reading nested fields. Use `--verbose` only when the default doesn\'t explain a failure \u2014 when you need `platform_transcript` (to check Vent\'s STT), per-turn or component-level latency breakdowns, the raw tool-call timeline, or provider-native artifacts in `debug.provider_metadata`. Otherwise skip \u2014 it just adds noise.\n\nVent\'s transcript is ground truth. Judge on semantic intent: ignore homophones and minor mis-hears (`"check teach hat"` for `"check that"`, missing question marks on short tails) \u2014 those are streaming-STT noise on Vent\'s caller side, not agent bugs, and **don\'t surface them in the report** (they\'re Vent-side artifacts, not actionable for the user). But clear gibberish or word-soup (e.g. `"Cristoxin"` where the agent should have said `"Of course, talk soon"`) is **not** a Vent artifact \u2014 Vent\'s STT does not invent words like that. It means the platform\'s TTS produced corrupted audio or the agent\'s STT/LLM generated the wrong text, and the fix lives there (TTS voice config, agent prompt, model temperature, codec issue). Missing or cut-off agent text, unlike word-soup, can come from Vent\'s STT stream dropping: check the recording before blaming the agent.\n\nFor transfers: `call_metadata.transfer_attempted` (provider claimed) and `transfer_completed` (from the platform\'s call data) can disagree \u2014 report both. `transfers[]` carries destination, type, and per-attempt status.\n\n### Saved history\n\nAfter every run, Vent writes the full result JSON to `.vent/runs/`. Shape:\n\n```jsonc\n{\n  "run_id": "...",\n  "timestamp": "2026-04-21T...Z",\n  "git_sha": "...",\n  "summary": { "calls_total": 2, "total_duration_ms": 12345, "total_cost_usd": 0.01 },\n  "call_results": [\n    { "name": "happy-path", "status": "completed", "duration_ms": 6123, "transcript": [], "observed_tool_calls": [], "latency": { "response_time_ms": 420, "p95_response_time_ms": 980 }, "call_metadata": { "cost_usd": 0.004 } }\n  ]\n}\n```\n\nWhen comparing against a prior run (Workflow step 6), inspect:\n\n- Run-completion status flips: `call_results[i].status` (pipeline-only \u2014 judge mission success from the transcript)\n- Latency: `call_results[i].latency.response_time_ms` (mean) or `latency.p95_response_time_ms` increased >20%\n- Tool calls: count of `call_results[i].observed_tool_calls[].successful` dropped\n- Cost: `summary.total_cost_usd` or `call_results[i].call_metadata.cost_usd` increased >30%\n- Transcript: `call_results[i].transcript` diverged in semantic content (ignore STT noise)\n\n## Reporting Results\n\nBefore reporting, read the agent\'s code to locate where the observed behavior originates. If the issue is small and you can fix it, fix it and explain what you did \u2014 don\'t ask permission first.\n\nAdapt the report shape to the call \u2014 a clean pass needs little, a regression with a multi-layer cause needs more. Use a transcript excerpt when it helps the user see what happened.\n\nHard rules:\n\n- Pair raw numbers with their plain-English meaning \u2014 don\'t drop the number, but don\'t leave it unexplained. E.g. "p95 latency was 850ms, which is snappy and well within natural conversational pacing" or "p95 hit 1.6 seconds with the LLM as the bottleneck \u2014 noticeably sluggish to a caller."\n- Name the user\'s voice agent by platform on first mention (e.g. "the Vapi agent responded snappily throughout") so the user knows immediately which agent the observation is about. After that, just say "the agent" \u2014 don\'t repeat the platform name on every line.\n- Always include the recording from `call_metadata.recording_url` as an inline `[Recording](url)` link, placed in **one block at the very end of the report** \u2014 never sprinkled through the prose. Single call: one link as the last line. Multi-call: one labeled link per call (e.g. `reschedule-appointment: [Recording](url)`). Never paste a bare URL.\n- Mission success is your judgment, not Vent\'s. The per-call `status` is only `"completed"` (pipeline ran) or `"error"` (pipeline failed); decide whether the agent actually accomplished the scenario from the transcript and tool calls.\n- Similar-sounding word substitutions (e.g. "ocean" for "OSHA") are STT ambiguity, not comprehension failure. The fix lives in STT keyword hints, custom vocabulary, or a prompt-level term list \u2014 not the agent\'s reasoning.\n- Surface only what the user can act on in their own agent\'s code or config \u2014 never Vent-side artifacts (caller wait modes, harness timing, internal pipeline quirks), or `cost_usd` unless asked.\n\nFor multi-call runs, lead with your own judgment of what happened across the calls (e.g. "3 of 4 did what they were supposed to; `cancel-appointment` never actually canceled"), not a parroted pass/fail count. Then cover each call with whatever depth it needs.\n';
+var windsurf_default = "## Windsurf Execution\n\nVent calls typically take 30 seconds to 2 minutes. Each `vent-hq run` is one shell tool call \u2014 wait for stdout (the JSON result) before responding. Don't background; use the JSON returned by `npx vent-hq run` directly. If Cascade's auto-execution level prompts for approval on `npx vent-hq` commands, ask the user to add `npx vent-hq *` to the workspace allow list once so subsequent calls flow without interruption.\n\nCascade runs shell tool calls in parallel within a turn \u2014 for multiple calls from one suite, issue each named call as its own separate shell tool call in the same turn (do not combine them with `&` and `wait`):\n\n```bash\nnpx vent-hq run -f .vent/suite.vapi.json --call happy-path\nnpx vent-hq run -f .vent/suite.vapi.json --call tool-path\n```\n\nStay within Cascade's per-turn parallel-tool-call budget \u2014 fan out at most ~6 calls in one turn and respect the provider concurrency caps below.\n";
+
+// src/skills/index.ts
+var DESCRIPTION = "Place real calls against the user's voice agent to verify recent code changes worked end-to-end. Trigger on test requests or after any non-trivial voice-agent change.";
+function assemble(parts) {
+  const skill = body_default.replace("{{EXECUTION}}\n", () => parts.execution).replace("{{LIVEKIT_NOTE}}", () => parts.liveKitNote ?? "").replace("{{PARALLEL_EXAMPLE}}", () => parts.parallelExample ?? "");
+  return parts.frontmatter ? `---
+${parts.frontmatter.join("\n")}
+---
+
+${skill}` : skill;
+}
+var claudeCodeSkill = assemble({
+  frontmatter: ["name: vent", `description: ${DESCRIPTION}`, "allowed-tools: Bash(npx vent-hq *)"],
+  execution: claude_code_default,
+  liveKitNote: " (See `Claude Code Execution` for how to compose the kill\u2192start\u2192wait\u2192submit as a single foreground command.)",
+  parallelExample: " with `&` and `wait`"
+});
+var cursorSkill = assemble({
+  frontmatter: [`description: ${DESCRIPTION}`, "alwaysApply: true"],
+  execution: cursor_default,
+  parallelExample: " with `&` and `wait`"
+});
+var codexSkill = assemble({ execution: codex_default });
+var windsurfSkill = assemble({
+  frontmatter: ["name: vent", `description: ${DESCRIPTION}`],
+  execution: windsurf_default
+});
 
 // src/lib/setup.ts
 var SUITE_SCAFFOLD = JSON.stringify(
@@ -5921,7 +5531,7 @@ var SUITE_SCAFFOLD = JSON.stringify(
     connection: {
       adapter: "websocket",
       start_command: "TODO: command to start your agent (e.g. npm run start)",
-      agent_port: 3001
+      agent_port: DEFAULT_AGENT_PORT
     },
     calls: {
       "happy-path": {
@@ -5936,19 +5546,19 @@ var SUITE_SCAFFOLD = JSON.stringify(
 async function installClaudeCode(cwd) {
   const dir = path3.join(cwd, ".claude", "skills", "vent");
   await fs5.mkdir(dir, { recursive: true });
-  await fs5.writeFile(path3.join(dir, "SKILL.md"), claude_code_default);
+  await fs5.writeFile(path3.join(dir, "SKILL.md"), claudeCodeSkill);
   printSuccess("Claude Code: .claude/skills/vent/SKILL.md", { force: true });
 }
 async function installCursor(cwd) {
   const dir = path3.join(cwd, ".cursor", "rules");
   await fs5.mkdir(dir, { recursive: true });
-  await fs5.writeFile(path3.join(dir, "vent.mdc"), cursor_default);
+  await fs5.writeFile(path3.join(dir, "vent.mdc"), cursorSkill);
   printSuccess("Cursor: .cursor/rules/vent.mdc", { force: true });
 }
 async function installWindsurf(cwd) {
   const dir = path3.join(cwd, ".windsurf", "skills", "vent");
   await fs5.mkdir(dir, { recursive: true });
-  await fs5.writeFile(path3.join(dir, "SKILL.md"), windsurf_default);
+  await fs5.writeFile(path3.join(dir, "SKILL.md"), windsurfSkill);
   printSuccess("Windsurf: .windsurf/skills/vent/SKILL.md", { force: true });
 }
 var VENT_MARKERS = [
@@ -5964,11 +5574,11 @@ async function installCodex(cwd) {
   }
   const markerIndex = VENT_MARKERS.map((marker) => existing.indexOf(marker)).filter((idx) => idx >= 0).sort((a, b) => a - b)[0];
   if (markerIndex != null) {
-    await fs5.writeFile(filePath, existing.slice(0, markerIndex).trimEnd() + "\n\n" + codex_default + "\n");
+    await fs5.writeFile(filePath, existing.slice(0, markerIndex).trimEnd() + "\n\n" + codexSkill + "\n");
   } else if (existing) {
-    await fs5.writeFile(filePath, existing.trimEnd() + "\n\n" + codex_default + "\n");
+    await fs5.writeFile(filePath, existing.trimEnd() + "\n\n" + codexSkill + "\n");
   } else {
-    await fs5.writeFile(filePath, codex_default + "\n");
+    await fs5.writeFile(filePath, codexSkill + "\n");
   }
   printSuccess("Codex: AGENTS.md", { force: true });
 }
@@ -6042,13 +5652,8 @@ async function initCommand() {
 
 // src/commands/stop.ts
 async function stopCommand(args) {
-  const accessToken = await loadAccessToken();
-  if (!accessToken) {
-    printError("Not authenticated. Run `npx vent-hq init` first.");
-    return 2;
-  }
   try {
-    const res = await apiFetch(`/runs/${args.runId}/stop`, accessToken, {
+    const res = await apiFetch(`/runs/${args.runId}/stop`, {
       method: "POST"
     });
     const data = await res.json();
@@ -6068,26 +5673,10 @@ async function stopCommand(args) {
 }
 
 // src/lib/dotenv.ts
-import { readFileSync } from "node:fs";
-import { resolve } from "node:path";
-function loadDotenv(dir = process.cwd()) {
+function loadDotenv() {
   for (const file of [".env.local", ".env"]) {
     try {
-      const content = readFileSync(resolve(dir, file), "utf-8");
-      for (const line of content.split("\n")) {
-        const trimmed = line.trim();
-        if (!trimmed || trimmed.startsWith("#")) continue;
-        const eqIdx = trimmed.indexOf("=");
-        if (eqIdx === -1) continue;
-        const key = trimmed.slice(0, eqIdx).trim();
-        let value = trimmed.slice(eqIdx + 1).trim();
-        if (value.startsWith('"') && value.endsWith('"') || value.startsWith("'") && value.endsWith("'")) {
-          value = value.slice(1, -1);
-        }
-        if (process.env[key] === void 0) {
-          process.env[key] = value;
-        }
-      }
+      process.loadEnvFile(file);
     } catch {
     }
   }
@@ -6104,17 +5693,21 @@ Commands:
   login        Authenticate via browser
   logout       Remove saved credentials
 Options:
-  --help    Show help
-  --version Show version
+  --help, -h     Show help
+  --version, -v  Show version
 
 Run 'npx vent-hq <command> --help' for command-specific help.`;
+var INIT_USAGE = `Usage: vent-hq init
+
+Authenticate (GitHub CLI token or browser sign-in), install the Vent skill
+files for Claude Code, Cursor, Codex and Windsurf, and scaffold
+.vent/suite.json if it does not exist.`;
 var RUN_USAGE = `Usage: vent-hq run -f <suite.json> [options]
 
 Options:
   --file, -f     Path to suite JSON file (required)
   --call         Name of the call to run (required if suite has multiple calls)
-  --session, -s  Reuse an existing local agent session
-  --verbose, -v  Include verbose fields in the result JSON`;
+  --session, -s  Reuse an existing local agent session`;
 var AGENT_USAGE = `Usage: vent-hq agent <command> [options]
 
 Commands:
@@ -6127,16 +5720,34 @@ Start options:
 
 Stop options:
   vent-hq agent stop <session-id>`;
+var STOP_USAGE = `Usage: vent-hq stop <run-id>
+
+Cancel a queued or running call.`;
+var LOGIN_USAGE = `Usage: vent-hq login [options]
+
+Options:
+  --status       Check whether you are logged in (exit code 1 if not)`;
+var LOGOUT_USAGE = `Usage: vent-hq logout
+
+Remove saved credentials from ~/.vent/credentials.`;
+var COMMAND_USAGE = {
+  init: INIT_USAGE,
+  run: RUN_USAGE,
+  agent: AGENT_USAGE,
+  stop: STOP_USAGE,
+  login: LOGIN_USAGE,
+  logout: LOGOUT_USAGE
+};
 async function main() {
   loadDotenv();
   const args = process.argv.slice(2);
   const command = args[0];
-  if (!command || command === "--help" || command === "-h") {
-    console.log(USAGE);
+  if (!command || args.includes("--help") || args.includes("-h")) {
+    console.log(COMMAND_USAGE[command ?? ""] ?? USAGE);
     return 0;
   }
   if (command === "--version" || command === "-v") {
-    const pkg = await import("./package-ZEOXQ7DY.mjs");
+    const pkg = await import("./package-WDXEE7G3.mjs");
     console.log(`vent-hq ${pkg.default.version}`);
     return 0;
   }
@@ -6146,17 +5757,12 @@ async function main() {
       return initCommand();
     }
     case "run": {
-      if (commandArgs.includes("--help")) {
-        console.log(RUN_USAGE);
-        return 0;
-      }
       const { values } = parseArgs({
         args: commandArgs,
         options: {
           file: { type: "string", short: "f" },
           call: { type: "string" },
-          session: { type: "string", short: "s" },
-          verbose: { type: "boolean", short: "v", default: false }
+          session: { type: "string", short: "s" }
         },
         strict: true
       });
@@ -6168,13 +5774,12 @@ async function main() {
       return runCommand({
         file: values.file,
         call: values.call,
-        session: values.session,
-        verbose: values.verbose
+        session: values.session
       });
     }
     case "agent": {
       const subcommand = commandArgs[0];
-      if (!subcommand || subcommand === "--help" || subcommand === "-h") {
+      if (!subcommand) {
         console.log(AGENT_USAGE);
         return 0;
       }
@@ -6206,9 +5811,9 @@ async function main() {
     }
     case "stop": {
       const runId = commandArgs[0];
-      if (!runId || commandArgs.includes("--help")) {
-        console.log("Usage: vent-hq stop <run-id>");
-        return runId ? 0 : 2;
+      if (!runId) {
+        console.log(STOP_USAGE);
+        return 2;
       }
       return stopCommand({ runId });
     }