/**
 * Transcription vocabulary bias. OpenAI's gpt-4o-transcribe accepts a `prompt` that
 * nudges the model toward expected words + spellings. Without it, homophones get
 * mis-heard out of context — a developer saying "bugs" comes back as "bars". We build
 * the prompt from what the agent IS (its capabilities) plus the proper nouns the platform
 * and the user can name, so each agent biases toward its own domain vocabulary.
 *
 * ── WHAT BELONGS IN THIS LIST (the rule, applied everywhere below)
 *
 * A term earns a place only if it is BOTH **likely to be said** to this agent and **likely to be
 * mis-heard** by a general-purpose recogniser. Both halves matter, and the second one is the one
 * that keeps the list short:
 *
 *   - `tmux`, `pnpm`, `ATS` — said constantly, and not in the decoder's ordinary English prior.
 *     IN.
 *   - the agent's own name, an attached repo's name — said constantly ("ask Heartfull to…"),
 *     and by construction not a word. IN.
 *   - `file`, `open`, `please`, `yes` — said constantly and transcribed correctly every time.
 *     OUT. They dilute the list and buy nothing.
 *   - a commit SHA, a file path, a collection id, a URL — knowable, and never said out loud in
 *     that form. OUT, even though we could enumerate them.
 *
 * The list is finite (`prompt` is a bounded field and dilution degrades bias), so every term
 * added is a term of budget spent — see MAX_EXTRA_TERMS.
 *
 * ── AND WHY A LONGER LIST IS NOT A FREE WIN (#332)
 *
 * The bias list is also a hazard. Given a low-information clip the decoder does not return
 * nothing: it returns a term from the list it was handed, and that text is indistinguishable
 * from the user's own words. `isTranscribeBiasEcho` below catches the multi-word case. It
 * deliberately does not catch single words, so every SINGLE-WORD term added here is a term that
 * can come back as a phantom turn uncaught. That is the direct cost of a longer list, and it is
 * why the rule above is "likely to be mis-heard" rather than "we happen to know it".
 */
/**
 * How many derived/user terms may ride along with the static lists.
 *
 * The `prompt` field is bounded and its bias quality DEGRADES with length — a 200-term list is
 * worse than a 20-term one, not better, because every term is a candidate the decoder can reach
 * for. 30 is roughly "the proper nouns one person says to one agent in a working session", which
 * is what the caller is asked to supply most-recently-used-first so the cap truncates the tail
 * rather than the head.
 */
export declare const MAX_EXTRA_TERMS = 30;
/** Split an identifier into the words it is SAID as: `tmux-operator-runner` is three words out
 *  loud, and `ProAgentStore/platform` is two. Returns "" when there is nothing to split, so the
 *  caller can skip adding a duplicate. */
export declare function spokenForm(term: string): string;
export interface TranscribePromptOptions {
    /**
     * The agent's declared `runtime` (`"coding"`, `"browser"`, …).
     *
     * #371's root cause was reading the wrong axis: the gate asked about SURFACES, a tmux operator
     * has `surfaces: ["tmux"]`, and so `CODING_TERMS` — which already contained `console`,
     * `terminal`, `git`, `npm`, `commit`, `deploy`, every word that user was saying — was skipped
     * for an agent whose `runtime` was `"coding"` the whole time. Both axes are read now: a
     * surface says what the agent SHOWS, a runtime says what it DRIVES, and either is enough to
     * know the user is talking about code.
     */
    runtime?: string | null;
}
/**
 * Build the transcription prompt for an instance from its capabilities (+ any extra proper
 * nouns: the agent's name, attached repo names, terminal targets — see #372). Returns "" when
 * there's nothing to bias, so the caller can omit the field entirely.
 */
export declare function buildTranscribePrompt(surfaces?: string[], extra?: string[], opts?: TranscribePromptOptions): string;
/**
 * Add proper nouns to an already-built prompt, applying the same dedupe/split/cap rules.
 *
 * Exists because the terms arrive from two places at two times: the console knows the agent's
 * capabilities at render, and the derived + user vocabulary (#372/#373) arrives with the voice
 * CONFIG, which `useVoice` refreshes on every mic start. Re-joining through here rather than
 * concatenating strings is what keeps the cap and the dedupe honest across both.
 */
export declare function extendTranscribePrompt(prompt: string, extra?: string[]): string;
/** Read a bias prompt back into its terms — the inverse of the join above. */
export declare function transcribeBiasTerms(prompt?: string): string[];
/**
 * Is this transcript just our own vocabulary list read back to us (#332)?
 *
 * The observed failure: two user messages in a row, both carrying audio, both reading **"Coder
 * Lead"** — the agent's own name, which the console adds to the bias list. The user said neither.
 * Given silence, the decoder does not return nothing; it returns a term from the list it was
 * handed, and a PROPER NOUN is the likeliest one to surface because it is the only distinctive
 * token among generic vocabulary. `365475c` fixed the fluent version of this (a prose framing
 * produced whole invented sentences) by making the prompt a bare list. A bare list still supplies
 * candidates — the failure got quieter, not fixed.
 *
 * ── Why MULTI-WORD terms only
 *
 * `isNoiseTranscript`'s whole contract is that genuine short commands survive: "yes", "no", "go",
 * "stop", "next" are deliberately NOT filtered. A one-word bias term is the same shape — "commit",
 * "deploy" and "refactor" are all things a user really does say alone to a coding agent, and
 * dropping one would trade a phantom turn for a swallowed one, which is the same bug wearing the
 * other hat. A multi-word term echoed back ALONE ("Coder Lead", "pull request") carries no
 * instruction at all, so the cost of asking for it again is close to zero.
 *
 * The rule is unchanged by #371/#372/#373, and it is the reason the term-entry rule at the top of
 * this file is written the way it is: a single-word addition is an addition this guard cannot
 * cover, so it has to earn its place on accuracy alone. Note that a derived or user-supplied
 * proper noun is usually helped here rather than hurt — `tmux-operator-runner` and
 * `ProAgentStore/platform` both arrive with a multi-word {@link spokenForm} beside them, and the
 * echo of EITHER form is caught.
 *
 * This is defence in depth, not the fix: silence should never reach the decoder in the first
 * place (see the dictation gate in `gate.ts` and `speechVerdict`). It is what catches the leak.
 */
export declare function isTranscribeBiasEcho(text: string, prompt?: string): boolean;
//# sourceMappingURL=prompt.d.ts.map