import { resolveEntryProviderKind } from "../providers/connection-resolution.js";
import { ROUTING_IDENTITY_PROVIDERS } from "../providers/inference/auth.js";
import {
  getCatalogProviderForModel,
  PROVIDER_CATALOG,
} from "../providers/model-catalog.js";
import { resolveCallSiteConfig } from "./llm-resolver.js";
import {
  type ContextWindow,
  DEFAULT_CONTEXT_WINDOW_MAX_INPUT_TOKENS,
  type LLMCallSite,
  type LLMConfig,
} from "./schemas/llm.js";
import type { ContextWindowConfig } from "./types.js";

export interface EffectiveContextWindow {
  provider: string;
  model: string;
  enabled: boolean;
  maxInputTokens: number;
  modelMaxInputTokens: number;
  defaultInputTokens: number;
  compactThreshold: number;
  summaryBudgetRatio: number;
  targetBudgetRatio: number;
  overflowRecovery: ContextWindow["overflowRecovery"];
  isLongContextEnabled: boolean;
  maxOutputTokens?: number;
}

export function resolveEffectiveContextWindow({
  llm,
  callSite,
  overrideProfile,
  forceOverrideProfile,
  selectionSeed,
}: {
  llm: LLMConfig;
  callSite: LLMCallSite;
  overrideProfile?: string;
  /**
   * Float `overrideProfile` above the call-site layers (see
   * `ResolveCallSiteOpts.forceOverrideProfile`). Threaded so context-window
   * sizing reflects the same profile the dispatch path resolves.
   */
  forceOverrideProfile?: boolean;
  /**
   * Per-conversation mix seed (the conversation id). Threaded so context-window
   * sizing for a mix profile reflects the same arm the dispatch path picks.
   */
  selectionSeed?: string;
}): EffectiveContextWindow {
  const resolved = resolveCallSiteConfig(callSite, llm, {
    overrideProfile,
    forceOverrideProfile,
    selectionSeed,
  });
  // Routing identities dispatch built-in catalog models, so the model's
  // catalog owner carries their limits. An entry-name provider resolves
  // through its row's kind instead: a custom-endpoint kind has no catalog
  // models, so its models keep the conservative default even when a model
  // id collides with a built-in one (the custom endpoint's "gpt-5.5" is not
  // OpenAI's). Labels with no row fall back to the model's catalog owner.
  const catalogProviderId = ROUTING_IDENTITY_PROVIDERS.has(resolved.provider)
    ? getCatalogProviderForModel(resolved.model)
    : PROVIDER_CATALOG.some((p) => p.id === resolved.provider)
      ? resolved.provider
      : (resolveEntryProviderKind(resolved.provider, resolved.model) ??
        getCatalogProviderForModel(resolved.model));
  const catalogModel = PROVIDER_CATALOG.find(
    (provider) => provider.id === catalogProviderId,
  )?.models.find((model) => model.id === resolved.model);

  const modelMaxInputTokens =
    catalogModel?.contextWindowTokens ??
    DEFAULT_CONTEXT_WINDOW_MAX_INPUT_TOKENS;
  const defaultInputTokens =
    catalogModel?.defaultContextWindowTokens ??
    DEFAULT_CONTEXT_WINDOW_MAX_INPUT_TOKENS;
  const maxInputTokens = Math.min(
    resolved.contextWindow.maxInputTokens,
    modelMaxInputTokens,
  );

  return {
    provider: resolved.provider,
    model: resolved.model,
    enabled: resolved.contextWindow.enabled,
    maxInputTokens,
    modelMaxInputTokens,
    defaultInputTokens,
    compactThreshold: resolved.contextWindow.compactThreshold,
    summaryBudgetRatio: resolved.contextWindow.summaryBudgetRatio,
    targetBudgetRatio: resolved.contextWindow.targetBudgetRatio,
    overflowRecovery: resolved.contextWindow.overflowRecovery,
    isLongContextEnabled: maxInputTokens > defaultInputTokens,
    maxOutputTokens: catalogModel?.maxOutputTokens,
  };
}

export function contextWindowConfigFromEffective(
  base: ContextWindowConfig,
  effective: EffectiveContextWindow,
): ContextWindowConfig {
  return {
    ...base,
    enabled: effective.enabled,
    maxInputTokens: effective.maxInputTokens,
    targetBudgetRatio: effective.targetBudgetRatio,
    compactThreshold: effective.compactThreshold,
    summaryBudgetRatio: effective.summaryBudgetRatio,
    overflowRecovery: effective.overflowRecovery,
  };
}