identity_for (prompts.py)

Everything the agent says, and everything that checks it said it. Two kinds of text live here and they are not the same kind of thing: FIXED TEXT — the disclosure and the probe. Spoken as close to verbatim as a speech-to-speech model allows, because both were demonstrably mangled

"""Everything the agent says, and everything that checks it said it.

Two kinds of text live here and they are not the same kind of thing:

  FIXED TEXT   — the disclosure and the probe. Spoken as close to verbatim as a
                 speech-to-speech model allows, because both were demonstrably
                 mangled when left to the model. The disclosure is a legal
                 requirement (§8); it cannot depend on winning a race against an
                 impatient human and line noise.
  INSTRUCTIONS — the system prompt. Judgment, in prose, for a model to apply.

The module is text, `os.getenv` and the standard library. It imports nothing from
LiveKit or the realtime stack: this is the part that changes after every call
that goes wrong, and it must be testable in a bare interp

... (truncated -- full source via MCP)