OmniRoute/open-sse/services/modelFamilyFallback.ts

/**
 * Model Family Fallback — Phase 2 Feature (T5)
 *
 * Implements two-phase model resolution:
 *   Phase 1 (static, pre-request): already done by model.ts alias resolution.
 *   Phase 2 (dynamic, post-error): when a provider returns a model-not-available
 *   error (400 with specific message or 404), we try sibling models within the
 *   same "family" before giving up.
 *
 * Inspired by Antigravity Manager's account-aware dynamic model remapping
 * (commit 6cea566, Mar 8 2026).
 */

import { getModelContextLimit } from "../../src/lib/modelCapabilities";
import { parseModel } from "./model.ts";
import { CONTEXT_OVERFLOW_REGEX, containsModelUnavailableMessage } from "./errorClassifier.ts";
import { getRegistryEntry } from "../config/providerRegistry.ts";

// ── Model Family Definitions ─────────────────────────────────────────────────

/**
 * Ordered candidate lists per model family.
 * First entry is the most preferred; fallback proceeds in order.
 */
const MODEL_FAMILIES: Record<string, string[]> = {
  // Gemini 3 / 3.1 Pro family — ordered by preference
  "gemini-3-pro": [
    "gemini-3.1-pro-preview",
    "gemini-3-pro-preview",
    "gemini-3.1-pro-high",
    "gemini-3-pro-high",
    "gemini-3.1-pro-low",
    "gemini-3-pro-low",
  ],
  "gemini-3.1-pro": [
    "gemini-3.1-pro-preview",
    "gemini-3-pro-preview",
    "gemini-3.1-pro-high",
    "gemini-3-pro-high",
    "gemini-3.1-pro-low",
    "gemini-3-pro-low",
  ],
  "gemini-3-pro-preview": [
    "gemini-3.1-pro-preview",
    "gemini-3-pro-high",
    "gemini-3.1-pro-high",
    "gemini-3-pro-low",
    "gemini-3.1-pro-low",
  ],
  "gemini-3.1-pro-preview": [
    "gemini-3-pro-preview",
    "gemini-3.1-pro-high",
    "gemini-3-pro-high",
    "gemini-3.1-pro-low",
    "gemini-3-pro-low",
  ],
  "gemini-3-pro-high": [
    "gemini-3.1-pro-high",
    "gemini-3-pro-preview",
    "gemini-3.1-pro-preview",
    "gemini-3-pro-low",
    "gemini-3.1-pro-low",
  ],
  "gemini-3.1-pro-high": [
    "gemini-3-pro-high",
    "gemini-3.1-pro-preview",
    "gemini-3-pro-preview",
    "gemini-3.1-pro-low",
    "gemini-3-pro-low",
  ],

  // Gemini 2.5 Pro family
  "gemini-2.5-pro": ["gemini-2.5-pro-preview-06-05", "gemini-2.5-pro-exp-03-25"],
  "gemini-2.5-pro-preview-06-05": ["gemini-2.5-pro", "gemini-2.5-pro-exp-03-25"],

  // Claude Mythos family (Fable 5) — flagship falls to the next-best Opus
  // tiers before the cheaper Sonnet, matching the Opus family ordering.
  "claude-fable-5": ["claude-opus-4-8", "claude-opus-4-7", "claude-sonnet-5"],

  // Claude Opus family
  "claude-opus-4-8": ["claude-opus-4-7", "claude-opus-4-6", "claude-sonnet-5"],
  "claude-opus-4-7": ["claude-opus-4-6", "claude-opus-4-5-20251101", "claude-sonnet-5"],
  "claude-opus-4-6": ["claude-opus-4-6-thinking", "claude-opus-4-5-20251101", "claude-sonnet-5"],
  "claude-opus-4-6-thinking": ["claude-opus-4-6", "claude-opus-4-5-20251101"],

  // Claude Sonnet family — Sonnet 5 is the newest tier; degrade to 4.6 → 4.5 → 4.
  "claude-sonnet-5": [
    "claude-sonnet-4-6",
    "claude-sonnet-4-5-20250929",
    "claude-sonnet-4-20250514",
  ],
  "claude-sonnet-4-6": ["claude-sonnet-4-5-20250929", "claude-sonnet-4-20250514"],
  "claude-sonnet-4-5-20250929": ["claude-sonnet-4-6", "claude-sonnet-4-20250514"],

  // GPT-5 family
  "gpt-5": ["gpt-5-mini", "gpt-4o"],
  "gpt-5.1": ["gpt-5.1-mini", "gpt-5", "gpt-4o"],
};

// ── Error Detection ──────────────────────────────────────────────────────────

/**
 * Error message fragments that indicate the requested model is unavailable
 * for the current account/provider, as opposed to a transient error.
 */
const MODEL_UNAVAILABLE_FRAGMENTS = [
  "model not found",
  "model_not_found",
  "model not available",
  "model is not available",
  "no such model",
  "unsupported model",
  "unknown model",
  "this model does not exist",
  "invalid model",
  "model not supported",
  "does not support",
  "not enabled for",
  "access to model",
  "improperly formed request", // Kiro 400 (model unavailable)
];

/**
 * Returns true if the HTTP status + error message indicates the model
 * itself is not available, not a transient server error.
 */
export function isModelUnavailableError(status: number, errorMessage: string): boolean {
  if (status === 404) return true;
  if (status !== 400 && status !== 403) return false;

  const msg = errorMessage.toLowerCase();
  if (MODEL_UNAVAILABLE_FRAGMENTS.some((fragment) => msg.includes(fragment))) return true;
  return containsModelUnavailableMessage(errorMessage);
}

export function isContextOverflowError(status: number, errorMessage: string): boolean {
  if (status !== 400) return false;
  return CONTEXT_OVERFLOW_REGEX.test(errorMessage);
}

// ── Fallback Resolution ──────────────────────────────────────────────────────

/**
 * Get the next fallback model from the same family.
 *
 * @param currentModel  The model that just failed
 * @param triedModels   Set of model IDs already tried (to avoid cycles)
 * @returns             Next model to try, or null if family exhausted
 */
export function getNextFamilyFallback(
  currentModel: string,
  triedModels: Set<string>
): string | null {
  const parsed = parseModel(currentModel);
  const bareModel = parsed.model || currentModel;
  const provider = parsed.provider || parsed.providerAlias || "";
  const prefix = provider ? `${provider}/` : "";

  // Normalize dots to hyphens so kiro/claude-opus-4.8 finds the right entry.
  // Fall back to the bare model name to support keys like "gemini-3.1-pro-high"
  // whose dots are part of the literal name, not a version separator.
  const lookupKey = bareModel.replace(/\./g, "-");
  const family = MODEL_FAMILIES[lookupKey] ?? MODEL_FAMILIES[bareModel];
  if (!family) return null;

  // Resolve the provider's supported model IDs so we can match notation (dot vs hyphen)
  const registryEntry = provider ? getRegistryEntry(provider) : null;
  const supportedIds = registryEntry ? new Set(registryEntry.models.map((m) => m.id)) : null;

  for (const candidate of family) {
    let resolvedCandidate = candidate;
    if (supportedIds && !supportedIds.has(candidate)) {
      // Try dot-notation variants: claude-opus-4-8 → claude-opus-4.8
      const dotVariant = candidate.replace(/-(\d+)-(\d+)$/, "-$1.$2");
      const dotVariant2 = candidate.replace(/-(\d+)-(\d+)-/, "-$1.$2-");
      if (supportedIds.has(dotVariant)) resolvedCandidate = dotVariant;
      else if (supportedIds.has(dotVariant2)) resolvedCandidate = dotVariant2;
    }
    const fullCandidate = `${prefix}${resolvedCandidate}`;
    if (!triedModels.has(fullCandidate)) {
      return fullCandidate;
    }
  }

  return null; // family exhausted
}

/**
 * Check if a model belongs to any registered family.
 */
export function isInModelFamily(model: string): boolean {
  const parsed = parseModel(model);
  const bareModel = parsed.model || model;
  return bareModel in MODEL_FAMILIES;
}

/**
 * Get all members of a model's family (including itself).
 */
export function getModelFamily(model: string): string[] {
  const parsed = parseModel(model);
  const bareModel = parsed.model || model;
  const prefix =
    parsed.provider || parsed.providerAlias ? `${parsed.provider || parsed.providerAlias}/` : "";

  const family = MODEL_FAMILIES[bareModel];
  if (!family) return [model];
  return [model, ...family.map((c) => `${prefix}${c}`)];
}

/**
 * Find a model with larger context window from a list of candidate models.
 * Uses models.dev synced capabilities to compare context limits.
 */
export function findLargerContextModel(
  currentModel: string,
  availableModels: string[]
): string | null {
  const currentParsed = parseModel(currentModel);
  const currentProvider = currentParsed.provider || currentParsed.providerAlias || "unknown";
  const currentModelId = currentParsed.model || currentModel;
  const currentLimit = getModelContextLimit(currentProvider, currentModelId) ?? 0;

  let bestModel: string | null = null;
  let bestLimit = currentLimit;

  for (const candidate of availableModels) {
    if (candidate === currentModel) continue;
    const parsed = parseModel(candidate);
    const provider = parsed.provider || parsed.providerAlias || "unknown";
    const modelId = parsed.model || candidate;
    const limit = getModelContextLimit(provider, modelId) ?? 0;

    if (limit > bestLimit) {
      bestLimit = limit;
      bestModel = candidate;
    }
  }

  return bestModel;
}