import type { ContextBundle } from '../context/types.js';
import type { Intent } from '../context/types.js';
import type { Category, Mode } from '../config/categories.js';
import type { UserProvidedSource } from './types.js';
/**
 * Single authoritative context assembly. All context sources merge here,
 * in priority order, so downstream strategies never have to juggle parallel
 * streams and readers can trace exactly why something was or wasn't included.
 */
export interface GroundingInputs {
    bundle?: ContextBundle;
    webSearchContext?: string;
    webSearchSources?: string[];
    platformInstructions?: string;
    platformHints?: string[];
    acceptedExamples?: AcceptedExample[];
    userProvidedSources?: UserProvidedSource[];
}
export interface AcceptedExample {
    originalPrompt: string;
    optimizedPrompt: string;
    category: Category;
    platform?: string;
    intent?: Intent;
    ts: number;
}
export interface GroundingOutput {
    /** Ready-to-inject user-prompt block. Empty string if nothing applies. */
    block: string;
    /** Which sources contributed, in order. Useful for trace. */
    sources: string[];
}
/**
 * Priority order (highest → lowest). Documented so contributors don't add
 * a 4th silo silently:
 *
 *   0. Caller-provided sources (ground_prompt explicit grounding — per-call)
 *   1. User pinned instructions (authoritative; the user said "always do X")
 *   2. Project rules (CLAUDE.md / AGENTS.md / .cursorrules / clarify.md)
 *   3. Active file context (what they're working on)
 *   4. Session: prior accepted examples for similar prompts (few-shot)
 *   5. Web search (if enabled; contextualizes open-ended requests)
 *   6. Workspace metadata (frameworks, languages, package name)
 *   7. Target model capability hints (for the downstream LLM's strengths)
 *   8. Custom platform instructions (.md file + inline)
 *   9. Built-in platform syntax hints
 */
export declare function buildGroundingContext(inputs: GroundingInputs): GroundingOutput;
/**
 * Reconcile user-supplied mode with the intent-derived recommended mode.
 * User-supplied wins unless it's undefined or the literal default — in which
 * case the analyzer's recommendation applies.
 */
export declare function reconcileMode(userMode: Mode | undefined, analyzerRecommended: Mode | undefined, userExplicitlyPassed: boolean): {
    mode: Mode;
    source: 'user' | 'analyzer' | 'default';
};
/**
 * Shape the LLM call to match the target model's capabilities. Small,
 * local models choke on 2KB system prompts and drown in high-max-token
 * budgets; large models benefit from richer system prompts.
 */
export interface PromptShape {
    systemPromptBudget: 'compact' | 'standard' | 'rich';
    maxTokens: number;
    temperature: number;
    includeExamples: boolean;
}
export declare function getPromptShape(bundle: ContextBundle | undefined, intent: Intent | undefined): PromptShape;