/**
 * Hardware profiles — pick the right tier ladder for the machine running Ollama.
 *
 * The same code runs against very different hardware. This module encodes
 * the "which model belongs on which tier for which box" decision explicitly
 * so it's a product choice, not an env-var scavenger hunt.
 *
 * - dev-rtx5080 (default): hermes3:8b ladder. Validated Hermes Agent
 *   integration path — Nous Research's hermes3:8b emits clean tool_calls
 *   over Ollama's /v1 chat endpoint and is the proven default for driving
 *   this MCP from an external agent (2026-04-19).
 * - dev-rtx5080-qwen3: Qwen 3 alternate rail. Same-family top-to-bottom
 *   Qwen 3 ladder for users who prefer Qwen tooling or want to compare.
 *   Picks up the `think` / THINK_BY_SHAPE plumbing in tiers.ts.
 * - m5-max: prod target. Qwen 3 ladder sized for 128GB unified memory.
 *
 * Per-tier env vars (INTERN_TIER_INSTANT, etc.) still override a profile's
 * model picks, so one-off experiments don't require a new profile.
 *
 * The retired `dev-rtx5080-llama` profile (Llama 3.1 on Deep) was dropped
 * at v2.0.0 — Llama 3.1 8B is obsolete, and the parity-rail experiment ran
 * its course.
 *
 * The retired qwen2.5 defaults (`qwen2.5:*-instruct-q4_K_M`) were dropped
 * at v2.0.0 — qwen2.5 is retired on modern Ollama installs and the
 * `INTERN_TIER_*` env knobs are sufficient for anyone pinning an older model.
 */
import type { Tier, TierConfig } from "./tiers.js";
export type ProfileName = "dev-rtx5080" | "dev-rtx5080-qwen3" | "m5-max";
export interface Profile {
    name: ProfileName;
    description: string;
    tiers: TierConfig;
    /**
     * Per-tier timeouts in ms. Lives on the profile (not as a global constant)
     * because cold-load behavior is hardware-bound — Instant needs 15s of
     * margin on a 16GB-VRAM dev box but only 5s on M5 Max unified memory.
     * Found via the first live dogfood pass: all Instant calls on RTX 5080
     * timed out at 5s before first token.
     */
    timeouts: Record<Tier, number>;
    /**
     * Tiers to prewarm on server startup. Targeted adoption aid for tiers
     * most likely to become habit (Instant), where cold-load drag would
     * poison early product feel. NOT a blanket "make everything hot" knob —
     * Workhorse and Deep are deliberately excluded so VRAM pressure and
     * unintended residency churn don't start mattering.
     */
    prewarm: Tier[];
}
export declare const PROFILES: Record<ProfileName, Profile>;
export declare const DEFAULT_PROFILE: ProfileName;
/**
 * Resolve the active profile from env. Selection order:
 *   1. INTERN_PROFILE env var, if a known name
 *   2. DEFAULT_PROFILE (dev-rtx5080) when INTERN_PROFILE is unset/empty
 *
 * If INTERN_PROFILE is set to a value that isn't a known profile, throws
 * CONFIG_INVALID with the available names. Silent fallback (prior behavior
 * through Stage A) would mask a typo'd profile against the wrong hardware
 * ladder and bury the signal in late tier-timeout errors.
 *
 * Per-tier env vars (INTERN_TIER_INSTANT, etc.) override the profile's picks.
 * Profile.timeouts are not env-overridable — they are a hardware property,
 * not a one-off tuning knob.
 */
export declare function loadProfile(env?: NodeJS.ProcessEnv): Profile;
/**
 * Validation regex for Ollama model identifiers (FT-002).
 *
 * Ollama tag identifiers follow `name[:tag]`:
 *   - `name` is lowercase letters / digits / dots / underscores / hyphens.
 *     `ollama pull` and the registry reject uppercase in this segment, so
 *     we enforce the strict form to catch typos early.
 *   - `tag` (after the colon) allows mixed case because real-world
 *     quantization labels routinely capitalize (e.g. `custom:model-q4_K_M`
 *     for K-quants, `phi3:mini-4k-instruct-q5_K_M`). Lowercasing the tag
 *     in the validator would false-positive on legitimate overrides.
 *
 * Spaces, slashes, and other punctuation are still rejected on both
 * sides — Ollama would refuse them at pull time anyway, and the typo
 * blast-radius (hours-later OLLAMA_MODEL_MISSING) is exactly what
 * FT-002 exists to prevent.
 *
 * The leading negative lookahead catches the `<name>-<digits><letter>`
 * dash-for-colon typo (`hermes3-8b`, `gpt-oss-120b`) where a `:` was
 * meant — the size-tag shape always carries a trailing letter (`8b`,
 * `120b`). It requires that trailing letter, so it does NOT fire on a
 * bare trailing version number (`glm-5`, `qwen-3`), which is a real,
 * colonless model id — narrowing a false-reject found in the 2026-07
 * health pass (L1). Colon-tagged ids (`…:8b`, `…:cloud`) are unaffected:
 * the lookahead is anchored to `$` and `:` is not in its character class.
 *
 * Exported so tests + the doctor CLI subcommand can reference the same
 * pattern (single source of truth for what counts as a model identifier).
 */
export declare const OLLAMA_MODEL_NAME_RE: RegExp;
/**
 * Validation bounds for Ollama `num_ctx` (FT-002).
 *
 * Lower bound 256 — Ollama refuses anything smaller as unworkable
 * (insufficient room for any meaningful prompt). Upper bound 1,048,576
 * (1M tokens) covers the largest currently-shipped open weights
 * (Llama 4 1M context, future Qwen 4) with headroom; anything beyond
 * is almost certainly a typo or stale config copied from a marketing
 * deck rather than an Ollama-supported value.
 */
export declare const NUM_CTX_MIN = 256;
export declare const NUM_CTX_MAX = 1048576;
/**
 * Validate an integer `num_ctx` value sourced from env or config.
 * No-op when unset/undefined. Throws `InternError('CONFIG_INVALID')`
 * when out of bounds or non-integer. Exported for use by future
 * env-override paths and by the doctor CLI subcommand.
 */
export declare function validateNumCtx(varName: string, value: unknown): void;
export interface EnvOverride {
    key: string;
    tier: Tier;
    from: string;
    to: string;
}
/**
 * Report tier-model env overrides relative to the active profile's baseline.
 *
 * Called ONCE at startup from main() so the operator sees, e.g.,
 *   "INTERN_TIER_DEEP overrides deep: hermes3:8b → custom:model-q4_K_M"
 * on stderr — instead of silently pinning the wrong model and wondering
 * later why benchmarks look off. Pure function, no side effects; caller
 * decides how to surface it.
 */
export declare function detectEnvOverrides(env?: NodeJS.ProcessEnv): EnvOverride[];
/** Resolved cloud configuration. Null when cloud is not opted into. */
export interface CloudConfig {
    /** Cloud base host, e.g. https://ollama.com. */
    host: string;
    /** Bearer API key (OLLAMA_API_KEY). */
    apiKey: string;
    /**
     * Cloud model per tier. instant/workhorse use INTERN_CLOUD_MODEL; deep uses
     * INTERN_CLOUD_DEEP_MODEL when set (else the same model). `embed` is carried
     * for shape only — embeddings never route to cloud.
     */
    tiers: TierConfig;
    /** Per-tier cloud-attempt timeout (ms). Cloud is far slower than local 8B. */
    timeouts: Record<Tier, number>;
    /** Context-window cap (tokens) applied to every cloud request (GPU-time control). */
    numCtx: number;
    /**
     * Standby mode (F2a, v2.9). True when a key is present but
     * OLLAMA_CLOUD_PRIMARY is not enabled: routing stays LOCAL-PRIMARY with
     * zero egress by default, and cloud serves only calls that explicitly
     * request `backend:'cloud'` (per-call escalation, RouteLLM-style
     * per-invocation routing). False = cloud-primary (v2.7.0 behavior).
     */
    standby: boolean;
}
/**
 * Resolve the cloud configuration from env. Three outcomes (F2a, v2.9):
 *   - OLLAMA_CLOUD_PRIMARY truthy (+ key) → cloud-PRIMARY config (standby:false)
 *   - key present, PRIMARY unset/falsy    → STANDBY config (standby:true) —
 *     local-primary, zero egress until a call requests backend:'cloud'
 *   - neither                             → null (local-only, zero egress)
 *
 * Throws CONFIG_INVALID (fail-fast at startup) when OLLAMA_CLOUD_PRIMARY
 * is enabled but OLLAMA_API_KEY is missing, or a cloud model name is malformed.
 *
 * Env surface:
 *   OLLAMA_CLOUD_PRIMARY        opt-in switch (1/true/yes/on)
 *   OLLAMA_API_KEY              bearer key (required when enabled)
 *   OLLAMA_CLOUD_HOST           default https://ollama.com
 *   INTERN_CLOUD_MODEL          default qwen3-coder-next:cloud (instant+workhorse+deep;
 *                               keep NON-thinking — see CLOUD_DEFAULT_MODEL)
 *   INTERN_CLOUD_DEEP_MODEL     optional deep-only override (e.g. glm-5:cloud,
 *                               deepseek-v4-pro:cloud — thinking OK here)
 *   INTERN_CLOUD_TIMEOUT_*_MS   per-tier cloud timeouts (instant/workhorse/deep)
 *   INTERN_CLOUD_NUM_CTX        cloud context-window cap (default 32768)
 */
export declare function loadCloudConfig(env?: NodeJS.ProcessEnv): CloudConfig | null;
/**
 * True when the cloud backend may serve a given call. The single decision
 * point the runner/chat/batch budget-summing sites share, so "standby stays
 * local" can never drift per call-site:
 *   - no cloud config          → false (local-only)
 *   - per-call backend:'local' → false (caller pinned this call local)
 *   - cloud-primary            → true
 *   - standby                  → only when the call explicitly requested
 *                                backend:'cloud'
 * Callers that cannot carry a per-call directive (batch, corpus-search
 * explain) call this without `backend` — under standby they stay local,
 * and their tier budgets must NOT be inflated by cloud timeouts.
 */
export declare function cloudMayServe(cloud: CloudConfig | null | undefined, backend?: "cloud" | "local"): boolean;
//# sourceMappingURL=profiles.d.ts.map