Voice Preflight

CALL-E speaks your task text to a person on the phone. You cannot hear it first. The platform exposes no voice or provider field, so the only way to know how a script sounds is to render the same text yourself.

This skill covers doing that with a provider you already pay for, then gating the call on what the audio and the text actually show.

When to use it

Before any call whose success depends on a specific string reaching the callee. That includes a one-time code a person has to read back, a callback number, a disclosure a rule requires you to speak, an amount, plus a reference somebody will repeat to you. Also use it whenever a task has been edited, because the failure this catches is a line disappearing during a tidy-up.

Skip it for a call whose script carries no such string. A general enquiry has nothing to survive verbatim.

The procedure

1. Declare what must survive. Write the script file with a locked entry per critical line, each with the exact text and a reason. The reason is quoted back in the refusal, so it travels with the failure instead of living in somebody's head.
2. Set the spoken budget. maxSpokenSeconds is your number, not a default. Pick the length past which the callee will hang up.
3. Preview first. preview reads no credential and contacts nothing. It shows what would be sent, where the credential would go and the offline findings.
4. Render and listen. render synthesises through your provider and measures the audio. Open the file. The tool tells you where to listen, it does not tell you what you will hear.
5. Gate on the exit code. 0 means nothing blocked, 20 means the script should not go out as written, 30 is a config or input error, 40 means the provider refused.

export ELEVENLABS_API_KEY="..."
npm run voice -- preview --script my-script.json --provider examples/provider.elevenlabs.json
npm run voice -- render  --script my-script.json --provider examples/provider.elevenlabs.json \
  --allow-host api.elevenlabs.io

Rules that are not negotiable

• The descriptor never holds a credential. It names the environment variable that does. Committing a key inside a descriptor is the one mistake this design exists to prevent, so the loader refuses an authEnv that is not shaped like a variable name, plus a static header carrying a long opaque value.
• Name the host before you send. There is no default trusted host, because this talks to no fixed vendor. https alone says the transport is encrypted and nothing about who is on the other end.
• Never treat a digit-run report as a diagnosis. It says a run of digits is present. Whether a provider reads it digit by digit is a question the audio answers.
• Never estimate a duration. When the container cannot be measured, the length check is skipped and the output says so. A number nobody measured is worse than an honest gap.
• This does not change the voice CALL-E uses. POST /v1/calls has no field for a voice or a provider, then it rejects unknown properties. Do not tell an operator otherwise.

See references/safety.md for what a script may and may not ask a person to say, and references/examples.md for three worked scripts.