/**
 * Tier shape and constants — tier→model selection lives in ./profiles.ts.
 *
 * Claude picks the tier by picking the tool. Tools declare their tier;
 * resolveTier() turns a Tier into the concrete model name from a TierConfig,
 * which comes from the active Profile.
 */
export type Tier = "instant" | "workhorse" | "deep" | "embed";
export interface TierConfig {
    instant: string;
    workhorse: string;
    deep: string;
    embed: string;
    /**
     * Per-tier `num_ctx` (Ollama context window) override (added v2.4.0).
     *
     * When a tier's `num_ctx` is set, the MCP server explicitly sends
     * `options.num_ctx = <value>` on every generate call routed to that
     * tier — both the initial attempt and any fallback. When a tier's
     * `num_ctx` is unset (or this whole map is undefined), the request
     * options block omits `num_ctx` entirely and Ollama falls back to its
     * model-loaded default. The "absent when unset" branch is what
     * preserves v2.3.0 backward-compat.
     *
     * Operational driver: hermes3:8b at 32K context on RTX 5080 16GB VRAM
     * spills to CPU and kills workhorse tool latency. The dev-rtx5080
     * profile lowers workhorse/instant to 8192/4096 so the model stays
     * resident in VRAM; deep remains unset so long-context briefs keep
     * current behavior. M5 Max has no spill problem (128GB unified) so all
     * tiers stay unset there.
     *
     * Scope guard: this is profile-level only. There is intentionally no
     * per-call `num_ctx` override on tool inputs in v2.4.0 — operators
     * tune by switching profiles or pinning a tier via env override.
     */
    num_ctx?: {
        instant?: number;
        workhorse?: number;
        deep?: number;
        embed?: number;
    };
}
export declare function resolveTier(tier: Tier, config: TierConfig): string;
/**
 * Resolve the per-tier `num_ctx` value for the active tier, if the profile
 * sets one. Returns undefined when the tier is unset — callers MUST then
 * omit the `num_ctx` field from the Ollama options block entirely, so
 * Ollama uses its model-loaded default. v2.3.0 backward-compat depends on
 * the "absent when unset" contract; never substitute a fake default value.
 */
export declare function resolveNumCtx(tier: Tier, config: TierConfig): number | undefined;
/**
 * Per-tier timeouts in ms. Used by guardrails/timeouts.ts to enforce
 * degradation rules and decide fallback behavior.
 */
export declare const TIER_TIMEOUT_MS: Record<Tier, number>;
/**
 * Fallback tier — what to degrade to when a tier's timeout fires.
 * Deep → workhorse → instant. Embed has no fallback (no cheaper embed tier).
 */
export declare const TIER_FALLBACK: Record<Tier, Tier | null>;
/**
 * Default temperatures by work shape.
 *
 * Calibrated for Qwen 3 (supported alongside default hermes3:8b). Qwen 3
 * has a documented regression vs Qwen 2.5: greedy decoding (temp 0)
 * degrades quality — the official model cards publish minimum-safe defaults,
 * which is why classify/extract/triage now floor at 0.2 instead of 0.1 and
 * never hit zero. hermes3:8b tolerates the higher floors without loss.
 *
 * Thinking vs non-thinking mode is controlled by THINK_BY_SHAPE below plus
 * the per-call `think` field on GenerateRequest. Non-thinking models
 * (hermes3:8b) ignore `think`, so the same table works on both ladders.
 *
 * Upstream reference: Qwen3 HF card recommends Temperature=0.7 / TopP=0.8
 * (non-thinking) and Temperature=0.6 / TopP=0.95 (thinking). The
 * structured-JSON shapes here stay cooler than the card defaults — small
 * models producing strict schema need less entropy — but never greedy.
 */
export declare const TEMPERATURE_BY_SHAPE: {
    readonly classify: 0.2;
    readonly extract: 0.2;
    readonly triage: 0.2;
    readonly summarize: 0.3;
    readonly research: 0.6;
    readonly draft: 0.6;
    readonly chat: 0.7;
};
/**
 * Sampler defaults keyed to Qwen3 official guidance. Apply alongside
 * TEMPERATURE_BY_SHAPE when the tool constructs its options block.
 * hermes3:8b applies these without issue.
 */
export declare const TOP_P_BY_MODE: {
    readonly non_thinking: 0.8;
    readonly thinking: 0.95;
};
/**
 * Thinking-mode by work shape.
 *
 * Load-bearing for Qwen 3 on Ollama. When `think=true` and the model is a
 * reasoning model, CoT content is emitted into the response's `thinking`
 * field AND consumes num_predict budget. For short-output tasks (classify,
 * extract, triage, short summaries) that budget is tight — a thinking
 * model can burn the entire num_predict on CoT and return an empty
 * `response`. The prompt-level `/no_think` soft-switch does NOT work on
 * Ollama — only the API field `think` does.
 *
 * Research / briefs / drafts get think=true: they benefit from CoT, and
 * their num_predict is sized for reasoning + response together.
 *
 * Non-thinking models (hermes3:8b) silently ignore the field, so declaring
 * it still costs nothing on the default ladder.
 */
export declare const THINK_BY_SHAPE: {
    readonly classify: false;
    readonly extract: false;
    readonly triage: false;
    readonly summarize: false;
    readonly research: true;
    readonly draft: false;
    readonly chat: false;
};
//# sourceMappingURL=tiers.d.ts.map