import type { Spec } from "@slowcook-ai/core";
/**
 * 0.11.16+ — bounded-attention spec slicing.
 *
 * Brew's per-iter prompt today embeds the entire spec body. For a story
 * with 12 invariants and 8 acceptance scenarios, that's ~2-4KB of
 * context the agent must consider when working on a single target test.
 * Most of it is irrelevant to the iteration's specific scope.
 *
 * This slicer derives a focused projection of the spec for a given
 * target test by keyword-matching identifiers extracted from the
 * test's title chain. Conservative: when the heuristic produces too
 * narrow a slice, falls back to the full spec rather than starve the
 * agent of context.
 *
 * Future (0.12.0+): testgen will tag each emitted test with its
 * `invariant_id`, and we'll resolve via that explicit mapping instead
 * of string matching. This 0.11.16 implementation is the heuristic
 * fallback that ships first.
 */
export interface SpecSlice {
    invariants: string[];
    acceptance_scenarios: string[];
    non_goals: string[];
    /** True when the slicer fell back to the full spec (heuristic too weak). */
    fellBack: boolean;
    /** For iter-log telemetry. */
    ratio: {
        invariants: {
            kept: number;
            total: number;
        };
        scenarios: {
            kept: number;
            total: number;
        };
    };
}
/**
 * Extract identifiers from a target test id like
 * `tests/integration/story-007-ui.test.tsx > BookmarksPage > clicking Save calls /api/bookmarks`.
 *
 * Returns a lowercased, deduped set of identifier-shaped tokens
 * (function names, route paths, common nouns). Filters stop-words.
 *
 * The set is what we'll grep invariants and scenarios against.
 */
export declare function extractIdentifiersFromTargetTest(targetTestId: string): Set<string>;
/**
 * Produce a focused slice of the spec for the given target test.
 *
 * Algorithm (0.11.17+ — top-K ranked):
 *
 *   1. Score each invariant + scenario by token overlap with identifiers
 *      extracted from the target test name.
 *   2. Sort by score descending.
 *   3. Keep top `keepInvariants` invariants and top `keepScenarios`
 *      scenarios (each defaulting to a small fixed number, NOT
 *      "everything with score > 0"). On focused stories where every
 *      invariant shares common terms, score>0 is almost always true,
 *      and slicing fails to narrow — top-K forces aggressive narrowing.
 *   4. Drop entries with score=0 even if they'd fit in the top-K
 *      (no relevance ≠ low relevance — exclude entirely).
 *   5. When fewer than `minKept` invariants have score > 0 AND there
 *      are more invariants total than we kept, return the FULL spec
 *      as a safety fallback (the agent doesn't have enough relevant
 *      context to work from).
 *
 * Validated 2026-04-25 on rewo's story-007 manifest: prior `score > 0`
 * algorithm produced ~97% retention (effectively no slicing) and 50%
 * fallback rate. Top-K with K=5 reduces retention to ~45% on the
 * same data while keeping the most-relevant invariants surfaced.
 */
export declare function sliceSpecForTarget(spec: Spec, targetTestId: string, options?: {
    /** Minimum kept invariants before falling back to full spec. */
    minKept?: number;
    /** How many top-scored invariants to keep. Default 5. */
    keepInvariants?: number;
    /** How many top-scored scenarios to keep. Default 4. */
    keepScenarios?: number;
}): SpecSlice;
/**
 * Render a slice as YAML-shaped prose for the brew prompt's "spec" block.
 * Mirrors the keys of a real spec but only emits non-empty arrays.
 */
export declare function renderSpecSlice(slice: SpecSlice, fullSpec: Spec): string;
//# sourceMappingURL=spec-slice.d.ts.map