ops-chatterbox

# Ops Chatterbox Use this skill for service-level Chatterbox diagnosis. Chatterbox is the speech renderer; caller skills own facts, identity, dream state, answer invariance, memory policy, and conversation logic. ## Commands ```bash skills/ops-chatterbox/run.sh doctor --json skills/ops-chatterbox/run.sh health --url http://127.0.0.1:8018/health --json skills/ops-chatterbox/run.sh container --name chatterbox-fork-agent-server --json skills/ops-chatterbox/run.sh render-smoke --live --text "Chatterbox smoke test." --json skills/ops-chatterbox/run.sh assess path/to/caller.py --json ``` All commands emit JSON receipts. A diagnostic command should return a receipt even when a service is unavailable; use `--strict` when a caller needs a non-zero exit for `ok:false`. ## Receipt Contract Every receipt includes: - `schema` - `skill` - `command` - `status` - `ok` - `live` - `mocked` - `checks` - `failures` - `next_actions` Live HTTP, Docker, and render probes must also include the target URL or container, elapsed time, and the redacted source of the evidence. Never print secret environment values; container diagnostics may list env key names and which keys were redacted. ## Boundary Rules - A `/health` payload that echoes a requested `voice_delivery` field is not proof that audio changed. - Per-render `affect_effect`, `pace_effect`, `tag_handling`, audio artifact metadata, and transcript or ASR readback are stronger evidence than request echo. - Keep `answer_text` separate from `tts_render_text` when adding sparse paralinguistic tags or nonverbal cues. - Do not pass raw user-supplied bracket or XML speech controls directly to TTS. - Treat Chatterbox-Turbo native event tags and base-engine emotion knobs as different channels unless a live receipt proves the selected backend applies both. For voice-agent policy, read `$best-practices-chatterbox-agent`; this ops skill does not replace that design contract. ## Common Misuse Run `assess <file>` before adding Chatterbox calls to another skill. It flags patterns such as: - treating `tag_handling.tags_interpreted` or request echo as acoustic proof; - passing `[laugh]`, `[sigh]`, or similar controls without tag/transcript gates; - using `finished_response_audio` without hashing or copying the audio artifact; - mutating `answer_text` instead of using a separate `tts_render_text`; - calling a live render path without preserving request and response JSON. ## Proof Boundary `sanity.sh` and the default agentic eval fixture prove deterministic CLI and receipt behavior only. They do not prove the local Chatterbox service is running, that a model is loaded, or that a rendered voice sounds emotionally different. Live claims require `health`, `doctor`, or `render-smoke --live` receipts from the target service. ## References (retrieve on demand — do not vendor) External docs drift; cite the canonical URLs and fetch them when needed with `/context7` (library docs) or `/fetcher` (any URL/PDF) rather than caching stale copies. Verified reachable (HTTP 200) 2026-08-24. - Chatterbox TTS (Resemble AI): <https://github.com/resemble-ai/chatterbox> - llms.txt (LLM-friendly doc index): <https://www.resemble.ai/llms.txt> ```bash skills/context7/run.sh "chatterbox tts resemble" skills/fetcher/run.sh "https://github.com/resemble-ai/chatterbox" ```