import { BrainSpec, type BrainSuppliesCreds } from 'rhachet/brains';
/**
 * .what = atom config type
 * .why = shared type for model configs
 */
export type BrainAtomConfig = {
    model: string;
    description: string;
    spec: BrainSpec;
};
/**
 * .what = credential keys required by fireworks ai
 * .why = enables type-safe credential lookup via rhachet's BrainSuppliesCreds
 */
export type CredsFireworks = {
    FIREWORKS_API_KEY: string;
};
/**
 * .what = supplies for fireworks brain supplier
 * .why = enables credential injection via keyrack shorthand or explicit getter
 *
 * .patterns:
 *   - keyrack shorthand: { keyrack: { owner: 'ehmpath', env: 'prod' } }
 *   - explicit getter: () => Promise<{ FIREWORKS_API_KEY: string }>
 */
export type BrainSuppliesFireworks = {
    creds: BrainSuppliesCreds<CredsFireworks>;
};
/**
 * .what = version-pinned fireworks ai atom slugs — the CANONICAL names
 * .why = enables a caller to pin an exact model version, so a provider's
 *        version churn cannot move the weights under them
 *
 * .shape = `fireworks/{family}/{tier}/{version}`
 *
 *         the tier is its own segment, so a pin and its generic are peers that
 *         differ by one word:
 *
 *           fireworks/deepseek/flash/v4.1     <- this exact version
 *           fireworks/deepseek/flash/latest   <- whichever version is current
 *
 *         it also lets tier tell two models apart where a suffix used to:
 *         `glm/pro/5.3` and `glm/flash/5.3` are one release in two tiers.
 *
 * .note = this union is the CANONICAL set, never the ACCEPTED set. a caller may
 *         also name a versionless or a pre-tier slug, and `slug/AtomSlug.ts`
 *         is the root that maps all three forms — read it first.
 *
 * .note = every slug below was verified by a LIVE chat completion on
 *         2026-09-22. a catalog read is not evidence of serve-ability — three
 *         ids that the catalog returned as available answered 404 on
 *         inference on 2026-09-16. only a real call proves a model serves.
 *
 * .note = slugs dropped on 2026-09-16, each for a live 404:
 *         - `fireworks/qwen/3.7-plus`   — `qwen3p7-plus` 404s; catalog lists it
 *         - `fireworks/minimax/2.7`     — `minimax-m2p7` 404s; catalog lists it
 *         - `fireworks/gpt-oss/20b`     — `gpt-oss-20b` 404s; absent from catalog
 *
 * .note = `fireworks/glm/5.1` was dropped earlier: pay-per-token for GLM-5.1 was
 *         deprecated effective 2026-08-07, so its id 404s on serverless. it
 *         remains on provisioned throughput only, which this package does not
 *         target. a deprecation is never fixable by a re-pin.
 *
 * .note = five of these carry a retirement (`RETIREMENT_BY_ATOM_SLUG`). they
 *         stay in this union and keep their config, because a live probe on
 *         2026-09-22 found that every one of them still serves. a retirement is
 *         an ANNOUNCEMENT, and to drop a slug on the announcement would break a
 *         caller whose model still answers.
 */
export type BrainAtomSlugFireworksPinned = 'fireworks/deepseek/pro/v4' | 'fireworks/deepseek/flash/v4.1' | 'fireworks/deepseek/flash/v4' | 'fireworks/kimi/pro/k3' | 'fireworks/kimi/code/k2.7' | 'fireworks/kimi/pro/k2.6' | 'fireworks/glm/pro/5.3' | 'fireworks/glm/flash/5.3' | 'fireworks/glm/pro/5.2' | 'fireworks/minimax/flash/m3' | 'fireworks/gpt-oss/flash/120b' | 'fireworks/nemotron/flash/3.5';
/**
 * .what = model configuration by slug
 * .why = maps slugs to api model names, descriptions, and specs
 *
 * .note = on `cache.get` — fireworks bills a cached prompt token at a REDUCED,
 *         per-model rate, never at zero. the discount is not uniform: it spans
 *         ~3% of the input rate (deepseek flash) to ~20% (kimi-k2.7-code,
 *         minimax-m3). so each model carries its own cited figure, and a `$0`
 *         here would be a placeholder, not a measurement.
 *
 * .note = on `cache.set` — fireworks publishes NO cache-write rate and its rate
 *         table holds no write column; a prompt cache fills as a side effect of
 *         the read path. so `set: $0` is verified, not a placeholder.
 *         `tokens.cache.set` stays 0 for the same reason: the api returns no
 *         write count to bill.
 *
 * .note = rates below are the STANDARD tier, which is what a plain serverless
 *         call receives. the Priority and Fast tiers cost more, and this
 *         package requests neither.
 *
 * .note = `context.tokens` is the exact `context_length` the models api
 *         reports, never a rounded vendor-page figure. so 1_048_576 rather
 *         than "1M", 262_144 rather than "256K".
 *
 * .note = `cost.time` values are ESTIMATES, not measurements. they order the
 *         models by rough throughput; they are not benchmarks.
 *
 * .note = `gain.cutoff` holds the fireworks deploy date, which is an UPPER
 *         bound on the model's knowledge cutoff. the true cutoff is
 *         unpublished for every model here, and a guess would read as a fact.
 *
 * .sources:
 *   - rates: https://docs.fireworks.ai/serverless/pricing (read 2026-09-16)
 *   - prompt cache: https://docs.fireworks.ai/guides/prompt-caching
 *   - models: https://api.fireworks.ai/inference/v1/models (read 2026-09-16)
 *   - api docs: https://docs.fireworks.ai/reference/chat-completions-1
 *   - benchmarks: https://benchlm.ai/benchmarks/sweVerified
 */
export declare const CONFIG_BY_ATOM_SLUG: Record<BrainAtomSlugFireworksPinned, BrainAtomConfig>;