AGENT SKILL

anvil-serving-voice-ops

Operate topology-owned anvil-serving voice audio and Realtime proxy lifecycle, Mini-to-Dark loopback forwarding, profiles, sidecars, and bounded voice benchmarks. Use for voice status, logs, start/stop/restart, OpenClaw Talk topology checks, and voice evidence that must remain separate from LLM serve qualification evidence.

License 许可
MIT
最近更新
2026-11-14
Stars
★ 1
Forks
0

01 前言 Front Matter

name
anvil-serving-voice-ops
description
Operate topology-owned anvil-serving voice audio and Realtime proxy lifecycle, Mini-to-Dark loopback forwarding, profiles, sidecars, and bounded voice benchmarks. Use for voice status, logs, start/stop/restart, OpenClaw Talk topology checks, and voice evidence that must remain separate from LLM serve qualification evidence.

03 SKILL.md 正文

Anvil Serving Voice Ops

Use this skill for voice-pipeline operations: validating sidecar manifests, rendering speech-to-speech commands, managing topology-owned STT/TTS, managing the persistent Mini Realtime proxy, selecting profiles, running the Mini-local forwarding bridge, and collecting bounded voice benchmarks.

Start Here

  1. Read README.md, CLAUDE.md, and docs/OPERATOR-SKILLS-AND-SUBAGENTS.md before changing behavior or reporting voice results as product evidence.
  2. Prefer voice_manage for structured STT/TTS lifecycle and voice_proxy_manage for persistent Realtime proxy lifecycle. Use the existing CLI verbs for foreground, profile, sidecar, bridge, and benchmark operations that do not have MCP wrappers: anvil-serving voice sidecar validate, anvil-serving voice sidecar command, anvil-serving voice sidecar compose, anvil-serving voice audio {up|down|status|logs}, anvil-serving voice profiles {list|validate}, anvil-serving voice proxy {run|up|down|restart|status|logs|bridge}, and anvil-serving voice benchmark.
  3. Use 127.0.0.1 for local URLs. Do not introduce localhost.
  4. Pass secrets by environment variable name only. Do not put literal API keys in manifests, commands, compose snippets, packets, or logs.
  5. Record the command host before interpreting loopback. In the reference OpenClaw topology, Companion Node runs Gateway plus Anvil Voice Realtime/proxy and reserves its 16 GB RAM for OpenClaw, Claude Code, and Codex. Do not run STT/TTS/LLM model serves on Mini for reference testing. Use dark-audio for Dark bridge ports or mini-dark-audio-proxy for Mini proxy ports forwarding to Dark. mini-audio is an explicit optional same-host/local-audio mode only. A non-gateway checkout cannot validate Mini-local proxy paths by calling its own 127.0.0.1.
  6. Treat voice benchmark output as voice-pipeline evidence. Voice results do not qualify an LLM serve or change a capability alias.
  7. Route multi-sample STT corpus qualification to skills/anvil-serving-stt-benchmark/SKILL.md. Keep this skill for single-turn audio/voice-pipeline evidence and topology lifecycle.
  8. Do not introduce one-off lifecycle, proxy, or port-forwarding scripts as the operational path. If repeatable voice operation needs a new control, it belongs in anvil-serving voice and, when agent-operated, MCP/controller.

Gates

Stop for a human gate before public or non-loopback Realtime binds, exposing a Realtime token, changing cloud/STT/TTS spend, installing or updating external speech-to-speech packages, or starting long-running foreground services on a shared host.

voice sidecar command --with-auth and voice sidecar compose --with-auth render environment variable references for authenticated router calls. The command path returns argv, so explicitly warn that the env-var reference appears in process argv. The compose path includes an auth exposure comment in the rendered service. Use either path only on private hosts where process and Docker metadata are protected.

Never use anvil-serving voice benchmark as a shortcut for LLM serve qualification. Voice benchmarks measure the voice pipeline: turn latency, TTFA, audio/STT/TTS behavior, realtime transport, and user-perceived loop quality. They can support a voice-pipeline status report, but an LLM serve still requires its own preflight, benchmark evidence, independent review, and human approval before an operator changes a capability alias.

Use workflow_packet_validate before handing a voice packet to the broader workbench. Voice artifacts must remain explicitly marked as non-promotion evidence.

Playbooks

  • Sidecar manifest validation: anvil-serving voice sidecar validate --config <manifest>.
  • Host command rendering: anvil-serving voice sidecar command --config <manifest> --json; add --with-auth only when the operator accepts the argv exposure warning.
  • Compose rendering: anvil-serving voice sidecar compose --config <manifest>; inspect loopback ports, host.docker.internal routing, image name, and auth comments.
  • Managed audio serves: prefer voice_manage for agent/controller operation. The CLI equivalents are anvil-serving voice audio up --topology <topology> --config <manifest> and voice audio down; preview mutations first and require --confirm for live execution. Use anvil-serving voice audio status and bounded anvil-serving voice audio logs for reads. External lifecycle serves should be reported as skipped rather than forced into local Docker control.
  • Profile selection: anvil-serving voice profiles list --config <manifest> to list profiles, anvil-serving voice profiles validate --config <manifest> --profile <name> before live work, then add --profile <name> to audio, proxy, or voice benchmark when switching audio topology. For reference OpenClaw Talk and candidate benchmarks, keep Mini model-free and select Dark-host audio or a Mini-side proxy to Dark. Use mini-audio only when explicitly testing the optional same-host/local-audio mode.
  • Mini-to-Dark audio bridge: anvil-serving voice proxy bridge --topology <topology> --config <manifest> --profile mini-dark-audio-proxy --dry-run first. Run it on Mini. Require topology-derived Dark targets and reject every non-loopback Mini listener.
  • Realtime proxy: prefer voice_proxy_manage for the persistent Mini process. The CLI equivalents are anvil-serving voice proxy up for persistent operation and anvil-serving voice proxy run for a foreground diagnostic. Use down/restart only after preview and explicit confirmation; use bounded status/logs for inspection. Treat unreachable STT/TTS/router endpoints as blockers. Never start audio lifecycle implicitly.
  • Bounded benchmark: anvil-serving voice benchmark --config <manifest> and capture JSON as voice-pipeline evidence with promoted=false. The benchmark is an endpoint client, not a Realtime-proxy lifecycle operation, so it has no proxy owner. Use --scope audio for an LLM-free TTS-to-STT loop. Keep audio topology and LLM selection separate: --profile selects the audio path and --candidate-overlay <toml> selects a candidate LLM configuration. Do not combine --candidate-overlay with direct candidate flags. A voice benchmark does not replace anvil-serving eval preflight, capacity benchmarking, or repeated anvil-serving eval benchmark quality evidence for LLM serve qualification. Publish a meaningful retained result through skills/anvil-serving-benchmark-docs/SKILL.md; keep measured hardware separate from protected/co-resident hardware and retain promoted=false. Apply its references/artifact-set-contract.md and anvil-serving.benchmark-artifact-set/v1 template around the native voice-benchmark-evidence/v1 file so workload, plan, identity, raw evidence, friction, restoration, and decision roles remain explicit. Use that contract's publication-ready result card and publication summary when the retained voice artifact is meaningful; a format-only refresh does not rerun the bounded voice benchmark.
  • Multi-sample STT qualification: use anvil-serving voice corpus prepare|validate and anvil-serving voice benchmark --scope stt through the canonical anvil-serving-stt-benchmark skill. Do not collapse its primary human set, separate synthetic set, repeated schedule, concurrency lane, or restoration checks into the bounded one-turn benchmark above.

Result Packet

Return an operator-workflow/v1 packet when coordinating with the broader workbench. Use artifacts entries with kind: "voice-benchmark" or kind: "voice-sidecar-render", evidence_scope: "voice-pipeline", and promotion_quality_evidence: false. Keep advisory_priors empty unless external voice-specific priors are explicitly identified as advisory-only. Keep promoted=false and human_gate_required=true for any workflow that asks for a capability-alias change, public binds, cloud enablement, or long-running service starts.

If the existing voice and voice sidecar verbs cannot cover the request, report the missing wrapper as a product gap after showing the closest safe existing command and why it is insufficient.

04 同仓库更多 Skills