Voice Agent scaffold

A complete program, not a fragment: it compiles as written. Copy it, rename the placeholder identifiers to your domain, and run axon check — the compiler will tell you what your renaming broke. An MCP client can also generate it through the axon.compose tool by naming the voice domain.

// AXON Voice Agent scaffold — audio-in / audio-out conversational agent
// with declared `ots:` codec transformations (μ-law 8kHz ↔ PCM16).
//
// This scaffold targets PSTN-style integrations (Twilio, Vonage) where
// inbound audio arrives μ-law-encoded and the LLM-streaming side
// expects PCM16. The ots declarations are first-class transformation
// surfaces; the runtime dispatches to native or ffmpeg backends.

// ── Types ─────────────────────────────────────────────────────────

type AudioFrame {
    samples:   Bytes
    sample_rate: Int
    codec:     Text
}

type Transcript { text: Text }
type Token      { piece: Text }

type VoiceCallRequest { frame: AudioFrame }

// ── Closed-catalogue OTS transformations ──────────────────────────

ots InboundMulawToPcm16 {
    teleology:       "Decode mu-law 8kHz inbound audio for LLM-streaming pipelines"
    homotopy_search: shallow
    loss_function:   "RMSE on reconstructed waveform"
}

ots OutboundPcm16ToMulaw {
    teleology:       "Encode PCM16 LLM output back to mu-law for the carrier"
    homotopy_search: shallow
    loss_function:   "RMSE on reconstructed waveform"
}

// ── Identity + grounding ──────────────────────────────────────────

persona VoiceConcierge {
    domain: ["voice-assistant", "phone-concierge"]
    tone: empathetic
    confidence_threshold: 0.7
    cite_sources: false
}

context VoiceCall {
    memory: session
    language: "en"
    depth: standard
    max_tokens: 1024
    temperature: 0.5
}

anchor SpeakClearly {
    require: source_citation
    confidence_floor: 0.6
    unknown_response: "Sorry, I didn't catch that. Could you repeat?"
    on_violation: log
}

// ── Tool — streaming backend ──────────────────────────────────────

tool VoiceBackend {
    // LLM-routed: the tool IS the model. No `provider:` (v2.69.0). `openai` was not a real provider slug.
    effects:  <network, stream:drop_oldest>
    timeout:  30s
}

// ── Streaming flow ────────────────────────────────────────────────

flow HandleVoiceFrame(frame: AudioFrame) -> Stream<Token> {
    step Transcribe {
        given: frame
        apply: InboundMulawToPcm16
        ask: "Transcribe the decoded audio frame."
        output: FlowEnvelope<Transcript>
    }
    step Respond {
        given: Transcribe.output
        apply: VoiceBackend
        ask: "Reply token-by-token in natural speech."
        output: Stream<Token>
    }
    return Respond.output
}

// ── HTTP boundary (SSE-back for streaming audio) ──────────────────

axonendpoint VoiceCallAPI {
    method:    post
    path:      "/v1/voice/call"
    body:      VoiceCallRequest
    execute:   HandleVoiceFrame
    output:    Stream<Token>
    backend:   auto
    transport: sse(axon)
    retries:   0
    timeout:   60s public: true
}